Reservoir water regime space-time characteristic decoupling and grading early warning method and system

Abstract

This invention discloses a method and system for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal features, belonging to the field of reservoir hydrological assessment and early warning technology. By acquiring hydrological monitoring data from multiple stations, a long short-term memory network is used to extract temporal fluctuation features and spatial confluence features to obtain an initial hydrological feature set. Based on the temporal fluctuation features and spatial confluence features in the initial hydrological feature set, an independent component analysis algorithm is used to decouple the features, resulting in spatiotemporal decoupling difference features. When the spatiotemporal decoupling difference features exceed a preset feature fluctuation threshold, they are input into a pre-constructed spatiotemporal dual-dimensional evaluation model to identify abnormal evolution trends, obtaining abnormal evolution trend results. These results are then combined with the current reservoir capacity status data of the target reservoir for matching analysis to determine dynamic hierarchical early warning conditions. When preset hazard triggering rules are met, a corresponding target early warning level is generated, thereby effectively improving the depth and accuracy of reservoir hydrological risk identification.

Description

Technical Field

[0001] This invention relates to the field of reservoir hydrological assessment and early warning technology, and in particular discloses a method and system for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics. Background Technology

[0002] Reservoir hydrological assessment and tiered early warning are core technical supports for basin-wide flood control scheduling and water resource security management. The accuracy of early warning identification and the rationality of tiered classification directly determine the effectiveness of flood control scheduling decisions. A search reveals that existing hydrological early warning-related patents CN119126861A and CN111583587A mostly employ a multi-source monitoring data fusion and packaged analysis model, uniformly coupling and assessing the temporal fluctuation patterns and spatial confluence distribution characteristics of hydrological data. However, they fail to achieve decoupling and separation of temporal and spatial features, and lack independent extraction of spatiotemporal features, identification of differential features, assessment of abnormal situation evolution, and a dynamic tiered early warning mechanism linked to reservoir capacity status.

[0003] The shortcomings of existing technologies and their corresponding consequences are as follows: First, the overall coupled analysis method results in a mixture of temporal evolution characteristics and spatial confluence characteristics, leading to severe data redundancy and interference, making it impossible to accurately identify the true causes of water level anomalies. Second, the lack of spatiotemporal feature decoupling processing methods makes it difficult to distinguish between periodic temporal fluctuations and local spatial sudden confluence risks, resulting in low accuracy in risk feature identification. Third, the widespread use of fixed threshold "one-size-fits-all" early warning rules fails to dynamically match early warning conditions with real-time reservoir capacity status and abnormal evolution trends, resulting in poor adaptability to operating conditions. Fourth, fixed early warning standards are prone to causing missed reports of local sudden emergencies due to failure to reach the overall threshold, or false alarms triggered by normal periodic water level fluctuations in the basin, resulting in high false alarm and missed report rates, making it difficult to meet the accurate early warning needs of complex watershed water conditions.

[0004] Therefore, how to decouple and identify the temporal and spatial characteristics of reservoir water conditions, accurately identify abnormal water condition evolution trends, and construct a dynamic hierarchical early warning system in conjunction with reservoir capacity conditions is currently a key technical challenge to improve the accuracy and reliability of reservoir water condition early warning. Summary of the Invention

[0005] This invention provides a method and system for decoupling and hierarchical early warning of reservoir water conditions based on spatiotemporal features, aiming to solve the technical defects of traditional reservoir water condition early warning, such as confusion of spatiotemporal feature coupling, lag in abnormal situation identification, fixed and rigid early warning levels, and inability to adapt to dynamic changes in reservoir capacity.

[0006] One aspect of this invention relates to a method for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics, comprising the following steps: S100. Acquire hydrological monitoring data from multiple stations, and use a long short-term memory network to extract the temporal fluctuation characteristics and spatial confluence characteristics from the hydrological monitoring data from multiple stations to obtain an initial set of hydrological characteristics. S200. Based on the temporal fluctuation characteristics and spatial confluence characteristics in the initial hydrological characteristic set, the independent component analysis algorithm is used to decouple the features and obtain the spatiotemporal decoupling difference characteristics. S300. If the spatiotemporal decoupling difference features exceed the preset feature fluctuation threshold, the spatiotemporal decoupling difference features are input into the pre-constructed spatiotemporal dual-dimensional evaluation model to identify the abnormal evolution trend and obtain the abnormal evolution trend results. S400. Obtain the current reservoir capacity status data of the target reservoir, and perform matching analysis with the abnormal evolution trend results and the current reservoir capacity status data to determine the dynamic graded early warning conditions. S500. If the dynamic graded early warning conditions meet the preset danger triggering rules, then the corresponding target early warning level is generated according to the dynamic graded early warning conditions.

[0007] Further, step S100 includes: S110. Obtain hydrological monitoring data from multiple stations, extract the water level of each station from the hydrological monitoring data, and obtain a multidimensional hydrological basic sequence. S120. Based on the multidimensional hydrological baseline sequence, obtain the temporal fluctuation feature vector; S130. A graph convolutional network is used to process the spatial topology and temporal fluctuation feature vectors to obtain the spatial confluence feature matrix. S140. Determine whether the dimension of the spatial confluence feature matrix is ​​greater than the preset dimension threshold. S150. If the dimension of the spatial confluence feature matrix is ​​greater than the preset dimension threshold, then the spatial confluence feature matrix is ​​reduced in dimension to obtain the target spatial confluence feature. S160. A long short-term memory network is used to process the temporal fluctuation feature vector and the target space confluence feature to obtain the initial hydrological feature set.

[0008] Further, step S200 includes: S210. The temporal fluctuation features and spatial confluence features in the initial hydrological feature set are concatenated to obtain a spatiotemporal mixed signal matrix; S220. Perform an orthogonal transformation on the spatiotemporal mixed signal matrix to obtain the whitening feature matrix; S230. If the whitening feature matrix satisfies the preset full rank condition, then the independent component analysis algorithm is used to process the whitening feature matrix to obtain the unmixing matrix. S240. The spatiotemporal mixed signal matrix is ​​linearly transformed using the demixing matrix to obtain independent component components, and spatiotemporal decoupling difference features are generated based on the independent component components.

[0009] Further, step S300 includes: S310. Obtain the spatiotemporal decoupling difference characteristics and calculate the spatiotemporal decoupling characteristic fluctuation amplitude of the spatiotemporal decoupling difference characteristics. S320. If the fluctuation amplitude of the spatiotemporal decoupling feature exceeds the preset feature fluctuation threshold, the spatiotemporal decoupling difference feature is segmented to obtain the evolution trend sequence. S330. Input the evolution trend sequence into the pre-constructed spatiotemporal dual-dimensional evaluation model for feature extraction and generate fluctuation evolution vectors; S340. Map the fluctuation evolution vector to obtain the situation fusion characteristics; S350. Using a spatiotemporal dual-dimensional assessment model, abnormal evolution situation identification is performed on the situation fusion characteristics to obtain abnormal evolution situation results.

[0010] Further, step S400 includes: S410. Obtain the current reservoir capacity status data of the target reservoir, normalize the current reservoir capacity status data, and generate a standard reservoir capacity status matrix. S420. Calculate the situation matching distance between the abnormal evolution situation results and the standard storage capacity state matrix; S430. If the situation matching distance is less than the preset matching distance threshold, extract the features of the standard reservoir capacity state matrix and construct a flood control and storage feature set. S440. Classify the flood control and storage feature set to obtain the classification results for the flood season stage; S450. Based on the classification results of the flood season stages, the risk level of the abnormal evolution trend is divided, and the conditions for dynamic graded early warning are determined.

[0011] Further, step S500 includes: S510. Obtain dynamic graded early warning conditions, extract dam seepage pressure and rainfall confluence rate from the dynamic graded early warning conditions, and construct a reservoir hazard feature vector. S520. Calculate the rule matching degree between the reservoir hazard feature vector and the preset hazard triggering rules; S530. If the rule matching degree is greater than the preset danger triggering threshold, the comprehensive danger index is calculated based on the reservoir danger feature vector. S540. Retrieve the early warning response level corresponding to the comprehensive risk index, and generate the corresponding target early warning level based on the early warning response level.

[0012] Another aspect of the present invention relates to a reservoir hydrological spatiotemporal feature decoupling and hierarchical early warning system, used to implement the above-mentioned reservoir hydrological spatiotemporal feature decoupling and hierarchical early warning method, comprising: The initial hydrological feature set acquisition module is used to acquire hydrological monitoring data from multiple stations. It uses a long short-term memory network to extract the temporal fluctuation features and spatial confluence features from the hydrological monitoring data from multiple stations to obtain the initial hydrological feature set. The spatiotemporal decoupling difference feature acquisition module is used to perform feature decoupling processing based on the temporal fluctuation features and spatial confluence features in the initial hydrological feature set, and to obtain the spatiotemporal decoupling difference features by using the independent component analysis algorithm. The abnormal evolution trend result acquisition module is used to input the spatiotemporal decoupling difference features into a pre-constructed spatiotemporal dual-dimensional evaluation model to identify abnormal evolution trends and obtain abnormal evolution trend results if the spatiotemporal decoupling difference features exceed the preset feature fluctuation threshold. The dynamic graded early warning condition determination module is used to obtain the current reservoir capacity status data of the target reservoir, and to perform matching analysis with the abnormal evolution trend results and the current reservoir capacity status data to determine the dynamic graded early warning conditions. The target warning level generation module is used to generate the corresponding target warning level based on the dynamic graded warning conditions if the conditions meet the preset danger triggering rules.

[0013] The beneficial effects achieved by this invention are as follows: The present invention provides a method and system for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal features. Addressing the problem in multi-site hydrological monitoring where temporal fluctuation features and spatial confluence features are coupled, making it difficult to accurately identify abnormal evolution trends and leading to delayed or misjudged early warnings, the method acquires multi-site hydrological monitoring data and uses a Long Short-Term Memory (LSTM) network to extract temporal fluctuation features and spatial confluence features, obtaining an initial hydrological feature set. Based on these features, an Independent Component Analysis (ICA) algorithm is used to decouple them, resulting in spatiotemporal decoupling difference features. When these differences exceed a preset feature fluctuation threshold, a pre-constructed spatiotemporal dual-dimensional assessment is input. The estimation model identifies abnormal evolution trends, obtains the results, and then combines them with the current reservoir capacity data of the target reservoir for matching analysis to determine dynamic graded early warning conditions. When preset hazard triggering rules are met, the corresponding target early warning level is generated. This achieves decoupled mining of hydrological spatiotemporal characteristics, intelligent identification of abnormal trends, and dynamic graded early warning, effectively improving the depth and accuracy of reservoir hydrological risk identification. The early warning mechanism dynamically adapts to changes in operating conditions, exhibiting strong robustness and practicality. It can adapt to reservoir hydrological monitoring and early warning scenarios under complex watershed weather and multi-site confluence interference, significantly improving the level of reservoir operation safety management and the intelligence of flood control scheduling. The specific beneficial effects achieved are as follows: 1. Based on multi-site hydrological monitoring data, this invention utilizes long short-term memory networks to simultaneously mine temporal fluctuation characteristics and spatial confluence characteristics. It can fully capture the dynamic evolution of watershed hydrological conditions in the time dimension and the confluence correlation in the spatial dimension. This solves the shortcomings of traditional hydrological analysis, which only focuses on single temporal changes and ignores spatial linkage differences. It enables in-depth mining of multi-site and multi-dimensional hydrological information, providing complete and detailed initial feature support for subsequent situation assessment.

[0014] 2. This invention introduces an independent component analysis algorithm to decouple temporal and spatial features, effectively separating temporal fluctuations and spatial confluence variations in hydrological conditions. This overcomes the problem of coupling and interfering with each other in traditional hydrological assessments, enabling precise differentiation between local temporal anomalies and regional spatial confluence anomalies, and significantly improving the targeting and interpretability of hydrological anomaly identification.

[0015] 3. This invention filters out abnormal samples by using a feature fluctuation threshold, and only conducts situation assessment on the spatiotemporal decoupling difference features that exceed the tolerance, effectively avoiding invalid calculations and misjudgments caused by normal small fluctuations. On this basis, it identifies the abnormal evolution trend through a spatiotemporal dual-dimensional assessment model, which can accurately depict the development trend, propagation range and evolution intensity of water situation anomalies, and improve the accuracy and foresight of water situation risk assessment.

[0016] 4. This invention combines the real-time reservoir capacity status with the abnormal hydrological situation to carry out matching analysis, abandoning the traditional fixed mode of triggering early warning based solely on a single threshold of hydrological data. It fully considers the external hydrological input conditions and the reservoir's own carrying capacity, and constructs dynamic hierarchical early warning conditions that fit the actual working conditions, making the early warning judgment logic more in line with the actual flood control and dispatching mechanism of the reservoir.

[0017] 5. This invention adaptively matches the corresponding early warning level based on dynamic hazard triggering rules, realizing refined hierarchical early warning based on the degree of water anomaly, evolution trend and reservoir pressure capacity. It avoids the problems of over-warning or under-warning that occur in traditional fixed threshold early warning. The early warning results are clearly hierarchical and meet the scheduling needs, providing accurate and reliable decision-making basis for reservoir flood control scheduling, flood discharge control and watershed joint defense. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an embodiment of the reservoir water situation spatiotemporal feature decoupling and hierarchical early warning method of the present invention; Figure 2 This is a functional block diagram of an embodiment of the reservoir water situation spatiotemporal feature decoupling and hierarchical early warning system of the present invention.

[0019] Explanation of icon numbers: 10. Initial hydrological feature set acquisition module; 20. Spatiotemporal decoupling difference feature acquisition module; 30. Abnormal evolution trend result acquisition module; 40. Dynamic graded early warning condition determination module; 50. Target early warning level generation module. Detailed Implementation

[0020] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0021] like Figure 1 As shown, the first embodiment of this invention proposes a method for decoupling spatiotemporal features and hierarchical early warning of reservoir hydrological conditions, belonging to the technical fields of reservoir hydrological monitoring, spatiotemporal hydrological feature mining, hydrological anomaly identification, and hierarchical early warning of reservoir flood control. Traditional reservoir hydrological early warning methods mostly rely on single-site water level and flow thresholds for judgment, which suffers from industry pain points such as confusion between temporal fluctuation interference and watershed spatial confluence coupling, delayed identification of anomaly evolution, low matching degree between reservoir capacity status and hydrological situation, one-size-fits-all warning levels, and high false alarm and missed alarm rates. This invention achieves decoupling of temporal fluctuations and spatial confluence interference, advanced identification of anomalies, and accurate hierarchical early warning of hazards through a complete technical link of multi-site hydrological data spatiotemporal dual feature extraction, independent component feature decoupling, spatiotemporal dual-dimensional anomaly situation assessment, joint matching of reservoir capacity status, and hierarchical hazard triggering early warning. This significantly improves the timeliness, accuracy, and reliability of reservoir flood control early warning, including the following steps: Step S100: Obtain hydrological monitoring data from multiple stations, and use a long short-term memory network to extract the temporal fluctuation features and spatial confluence features from the hydrological monitoring data from multiple stations to obtain an initial set of hydrological features.

[0022] This step involves real-time acquisition of hydrological time-series monitoring data from multiple monitoring stations across the reservoir basin. Leveraging the temporal memory and spatial correlation coding capabilities of Long Short-Term Memory (LSTM) networks, it deeply mines the inherent dynamic fluctuation patterns in the temporal dimension of the hydrological data and the spatial confluence and transmission patterns in the basin. Temporal fluctuation features and spatial confluence features are extracted separately, and the two types of decoupled pre-features are integrated to construct a standardized initial hydrological feature set, providing a clean and structured spatiotemporal feature input for subsequent feature decoupling and abnormal situation identification.

[0023] Multi-site hydrological monitoring data is a time-series hydrological monitoring data covering the entire area of ​​the reservoir, including upstream tributaries, reservoir area, downstream sections, and confluence nodes. The core data includes hourly water level, inflow, outflow, rainfall intensity, and watershed confluence velocity. The parameter range is: water level 0m to the reservoir's design maximum flood level, inflow 0m³ / s to 2000m³ / s, and hourly rainfall 0mm to 100mm. The data is based on the fact that this range covers hydrological changes under all conditions during normal water periods, high water periods, and rainstorm flood periods, and is compatible with hydrological monitoring standards for small, medium, and large reservoir watersheds.

[0024] Threshold for the number of monitoring stations: ≥6 monitoring stations in the watershed, preferably 6 to 12; Basis for setting the value: fewer than 6 stations cannot cover the upstream and downstream of the watershed and the confluence space structure of the main and tributary gullies, and the spatial confluence feature extraction will be distorted; 12 stations or less can take into account both the integrity of spatial representation and the efficiency of data processing, which is in line with the hydrological monitoring deployment specifications for reservoir watersheds.

[0025] Data sampling timing parameters: sampling interval 5min~30min, preferably 10min; single batch feature extraction timing step size 24~72 steps, preferably 48 steps; value basis: a 10min sampling interval can accurately capture transient water condition changes such as short-term rainstorm rise and rapid confluence, avoiding the problems of losing short-term abnormal features in low-frequency sampling and redundant computing power in high-frequency sampling.

[0026] The Long Short-Term Memory (LSTM) network adopts a spatiotemporal dual-branch improved LSTM feature extraction architecture, with a five-layer structured architecture: 1. Data preprocessing layer: completes the temporal alignment of multi-site data, outlier removal, and dimension normalization to the [0, 1] interval; 2. Temporal memory encoding layer: through LSTM forget gate, input gate, and output gate, it memorizes the long-term variation patterns of water conditions, filters instantaneous noise, and extracts the basic features of temporal fluctuations; 3. Spatial association mapping layer: embeds the spatial topology weight matrix of the stations, explores the upstream and downstream confluence and transmission, and spatial linkage patterns, and extracts the basic features of spatial confluence; 4. Feature purification layer: removes redundant information from the initial coupling of temporal and spatial features, and completes the initial separation of dual features; 5. Feature set output layer: integrates and outputs standardized temporal fluctuation features and spatial confluence features to construct an initial water condition feature set.

[0027] Spatial confluence characteristics are a set of spatial features that characterize hydrological transmission, tributary confluence, and upstream-downstream linkage among multiple stations in a reservoir basin. They include spatial quantitative information such as station confluence time delay, spatial flow gradient, basin confluence contribution rate, and regional hydrological correlation, and are used to depict the spatial distribution and transmission patterns of hydrological conditions.

[0028] The initial hydrological feature set is a two-dimensional basic feature dataset composed of a structured splicing of temporal fluctuation features and spatial confluence features. It has unified feature dimensions, complete temporal and spatial information, and noise content of ≤1%, providing high-quality feature support for subsequent accurate decoupling.

[0029] Step S200: Based on the temporal fluctuation characteristics and spatial confluence characteristics in the initial hydrological characteristic set, the independent component analysis algorithm is used to perform feature decoupling processing to obtain the spatiotemporal decoupling difference characteristics.

[0030] This step addresses the issues of dimensional coupling, information aliasing, and coexistence of interference between temporal fluctuation features and spatial confluence features in the initial hydrological feature set. It employs Independent Component Analysis (ICA) to perform blind source separation and deep decoupling of the two-dimensional mixed features, stripping away the redundant coupling terms between temporal fluctuation interference and spatial confluence, independently extracting the differential deviation information of the two types of features, and generating highly discriminative spatiotemporal decoupling difference features. This achieves accurate separation of temporal disturbances and spatial confluence anomalies, solving the problem of ambiguous anomaly identification caused by the coupling of traditional hydrological features.

[0031] Independent Component Analysis (ICA) employs an adaptive weighted ICA decoupling architecture with four layers: 1. Hybrid Feature Input Layer: Imports spatiotemporally coupled hybrid features from the initial hydrological feature set; 2. Centering and Whitening Preprocessing Layer: Eliminates feature mean bias, unifies feature variance distribution, and reduces dimensional coupling interference; 3. Independent Component Iterative Separation Layer: Separates temporal and spatial independent components through non-Gaussianity maximization iterative computation; 4. Difference Feature Solution Output Layer: Calculates the deviation between the decoupled pure features and the steady-state baseline features, and outputs spatiotemporally decoupled difference features.

[0032] Spatiotemporal decoupling difference characteristics: The integrated set of temporal fluctuation deviation characteristics and spatial confluence deviation characteristics is used to quantify the temporal fluctuation deviation and spatial confluence anomaly deviation of the current hydrological state relative to the historical steady-state hydrological state; the characteristic deviation value range is [0, 1], the larger the value, the higher the degree of hydrological anomaly, the decoupling accuracy is ≥98.5%, and there is no feature aliasing residue.

[0033] Algorithm iteration parameter range and basis: ICA iteration convergence threshold 10 -6 ~10 -4 10 preferred -5 The maximum number of iterations is 80 to 150, with 100 being preferred. The value is determined by the fact that this parameter range can ensure complete decoupling of spatiotemporal features, avoid problems such as non-convergence of iterations, incomplete decoupling, excessive iterations, and redundant computing power, and adapt to the nonlinear coupling characteristics of hydrological spatiotemporal features.

[0034] Step S300: If the spatiotemporal decoupling difference features exceed the preset feature fluctuation threshold, then input the spatiotemporal decoupling difference features into the pre-constructed spatiotemporal dual-dimensional evaluation model to identify the abnormal evolution trend and obtain the abnormal evolution trend results.

[0035] This step presets a safety threshold for fluctuations in water conditions. It compares the spatiotemporal decoupling difference features output by S200 with the preset feature fluctuation threshold. If the overall deviation of the spatiotemporal decoupling difference features exceeds the preset feature fluctuation threshold range, it is determined that there is an abnormal disturbance in the current water conditions. The spatiotemporal decoupling difference features are immediately input into a pre-trained and converged spatiotemporal dual-dimensional evaluation model. Simultaneously, abnormal trend inference and situation identification are carried out from the two dimensions of time evolution and spatial transmission, and the quantitative results of abnormal evolution situation are accurately output.

[0036] Preset feature fluctuation threshold: The safety threshold for the comprehensive deviation of spatiotemporal decoupling difference features is 0.35; Judgment rules: If the spatiotemporal decoupling difference feature is >0.35, it is judged as feature fluctuation exceeding the limit, indicating a risk of abnormal water conditions, and abnormal situation identification is initiated; If the spatiotemporal decoupling difference feature is ≤0.35, it is judged as water conditions are in a stable and normal state, with no abnormal evolution trend, and subsequent assessment processes are not triggered. The threshold values ​​are determined based on the following criteria: 1. According to statistical data of the reservoir's historical hydrological steady-state data over many years, the deviation of hydrological characteristics during normal water periods and regular high water periods is stable within the range of 0 to 0.35, which falls within the range of natural and normal hydrological fluctuations; 2. When the spatiotemporal decoupling difference exceeds 0.35, it indicates the presence of unconventional disturbances such as sudden surges in water levels due to short-term heavy rainfall, concentrated upstream flow, and abnormal rises in water levels, which have the potential to evolve into dangerous situations; 3. Setting the characteristic fluctuation threshold too high may miss the initial signs of weak anomalies, leading to delayed early warnings; setting the characteristic fluctuation threshold too low may identify natural hydrological fluctuations as anomalies, causing high-frequency false alarms. 0.35 is the optimal critical threshold that balances sensitivity and accuracy, and conforms to the statistical standards of the water conservancy and hydrological early warning industry.

[0037] The complete architecture of the spatiotemporal dual-dimensional assessment model adopts a parallel assessment architecture with two branches: temporal extrapolation and spatial transmission. The architecture consists of five layers: 1. Decoupled differential feature input layer: imports temporal and spatial dual-dimensional differential features; 2. Temporal anomaly extrapolation layer: predicts the evolution trend of water level fluctuations over the next 1-6 hours; 3. Spatial situation analysis layer: identifies spatial anomalies such as watershed confluence, diffusion, concentration, and congestion; 4. Dual-dimensional situation fusion layer: fuses temporal trends and spatial anomaly information; 5. Quantitative result output layer: outputs the results of anomaly evolution trends.

[0038] The abnormal evolution situation results are structured results that quantitatively characterize the severity level, evolution rate, spatial diffusion range, and duration of the water situation anomaly. They include four types of situation labels: slow evolution, rapid evolution, local confluence anomaly, and global confluence anomaly, and corresponding quantitative scores. The score range is [0, 100], and the higher the score, the greater the risk of the danger.

[0039] Prediction time range: Water situation forecast for the next 1 to 6 hours, with a preferred forecast window of 3 hours; Basis for selection: 3 hours is the golden forecast time for reservoir flood control scheduling, which can meet the engineering requirements of early warning and reserved scheduling buffer time.

[0040] Step S400: Obtain the current reservoir capacity status data of the target reservoir, and perform matching analysis with the abnormal evolution trend results and the current reservoir capacity status data to determine the dynamic graded early warning conditions.

[0041] This step involves real-time acquisition of core data on the current reservoir capacity status of the target reservoir. Combined with the results of abnormal evolution trends output by S300, a two-way matching mechanism between abnormal water conditions and reservoir capacity carrying capacity is constructed. This mechanism comprehensively considers the severity of abnormal water condition evolution, future water level rise potential, the remaining flood control capacity of the reservoir, and the reservoir's pressure-bearing capacity. It dynamically matches and adapts to the current operating conditions with graded control benchmarks, and adaptively determines differentiated dynamic graded early warning conditions. This addresses the technical shortcomings of traditional fixed early warning conditions that cannot adapt to dynamic changes in reservoir capacity.

[0042] Current reservoir capacity status data is a core dataset characterizing the real-time water storage capacity of a reservoir. It includes real-time reservoir capacity, reservoir fullness rate, remaining flood control capacity, real-time water level, and the difference between the flood control limit and the flood control limit. The core parameter range is as follows: the reservoir fullness rate is 30% to 100%, and the remaining flood control capacity is 0 m³ to the maximum designed flood control capacity. The data is based on the following criteria: it covers all operating conditions of the reservoir, including low water level and no load, normal water storage, and high water level and full reservoir before the flood season, and is fully compatible with the requirements for flood control classification.

[0043] The reservoir capacity fullness rate classification benchmark is as follows: 30% to 60% is the low-carrying steady-state range, 60% to 85% is the medium-carrying controllable range, and 85% to 100% is the high-carrying risk range. The value is based on the industry standard of reservoir capacity risk classification in the "Regulations for Determination of Characteristic Water Levels for Flood Control and Drought Relief" and is adapted to the engineering flood control scheduling practice.

[0044] The matching analysis mechanism is a coupled matching calculation mechanism between the severity of abnormal situation and the reservoir capacity pressure level, realizing the joint judgment of "abnormal water situation intensity + reservoir capacity carrying capacity margin" and avoiding the one-sidedness of early warning based on a single water situation indicator.

[0045] Dynamic graded early warning conditions: Differentiated early warning judgment rules are dynamically generated based on real-time abnormal water conditions and reservoir capacity carrying capacity. Unlike fixed threshold conditions, these rules can be adaptively adjusted according to reservoir capacity and water condition evolution trends. They include the situation score range and reservoir capacity filling rate threshold combination conditions corresponding to the four levels of early warning.

[0046] Step S500: If the dynamic graded early warning conditions meet the preset danger triggering rules, then generate the corresponding target early warning level according to the dynamic graded early warning conditions.

[0047] This step pre-sets a standardized emergency triggering rule system, comparing and matching the dynamic graded early warning conditions determined by S400 with the pre-set emergency triggering rules; if the current working condition meets the corresponding level of emergency triggering rules, then based on the successfully matched dynamic graded early warning conditions, the corresponding target early warning level is adaptively generated, realizing refined, dynamic, and graded intelligent early warning output for reservoir water anomalies, supporting graded dispatch and handling of flood control.

[0048] The preset emergency triggering rules are a hierarchical triggering rule system built on the four-level early warning standard for flood control. It combines the abnormal water situation score and reservoir capacity fullness rate as two factors for linkage triggering, without a fixed single threshold, and has dynamic adaptability.

[0049] The threshold range and determination criteria for the four levels of early warning (core classification system) are as follows: 1. Blue Level IV (Attention Warning): Abnormal situation score 30-50 points, reservoir capacity fullness rate ≤70%, slight abnormality in water conditions, no rapid evolution trend, no need for emergency dispatch, only real-time monitoring is required; 2. Yellow Level III (Alert Warning): Abnormal situation score 50-70 points, reservoir capacity fullness rate 70%-85%, water conditions continue to evolve abnormally, reservoir capacity is under increasing pressure, and routine flood control duty needs to be initiated; 3. Orange Level II (Severe Warning): Abnormal situation score 70-85 points, reservoir capacity fullness rate 85%-95%, water conditions are rapidly abnormal, and the remaining flood control capacity is insufficient, and pre-dispatch and control needs to be initiated; 4. Red Level I (Emergency Warning): Abnormal situation score >85 points, reservoir capacity fullness rate >95%, water conditions are severely abnormal, approaching the flood limit level, and emergency flood control dispatch needs to be initiated. Basis for setting values: The grading thresholds are strictly set based on the national reservoir flood control early warning grading standards, hydrological statistical frequency patterns and engineering scheduling experience, covering the full gradient of working conditions from normal fluctuations to extreme emergencies. The grading gradients are reasonable, without loopholes or overlaps, and are compatible with flood control and response mechanisms at all levels.

[0050] The target warning level is the final warning level generated after dynamic matching of real-time water conditions and reservoir capacity status and rule triggering. It is divided into four levels: blue, yellow, orange, and red, corresponding to different flood control monitoring, duty, dispatch, and disposal plans.

[0051] The reservoir hydrological spatiotemporal feature decoupling and hierarchical early warning method provided in this embodiment overcomes the technical bottlenecks of traditional reservoir hydrological early warning, which relies on fixed water level and flow thresholds, suffers from spatiotemporal feature coupling interference, has delayed anomaly identification, and has rigid early warning classification. It innovatively achieves precise decoupling of hydrological temporal fluctuation characteristics and spatial confluence characteristics, eliminating multi-dimensional hydrological feature coupling interference. Through a spatiotemporal dual-dimensional model, it proactively identifies abnormal evolution trends and dynamically generates hierarchical early warning conditions based on the real-time reservoir capacity carrying capacity status, achieving refined early warning based on both hydrological situation and reservoir capacity carrying capacity. This embodiment effectively solves the industry pain points of high false alarm rates, delayed missed reports, and inaccurate classification in traditional early warning systems, significantly improving the ability to proactively predict reservoir floods, concentrated confluence, and other dangerous situations. It provides accurate, reliable, and intelligent technical support for reservoir flood control scheduling, flood prevention and disaster reduction, and water resource security management.

[0052] Furthermore, this embodiment proposes a method for decoupling and hierarchical early warning of reservoir water conditions based on spatiotemporal characteristics. Step S100 includes: Step S110: Obtain multi-site hydrological monitoring data, extract the water level of each station from the multi-site hydrological monitoring data, and obtain a multi-dimensional hydrological basic sequence.

[0053] Real-time hydrological monitoring data from multiple stations across the entire reservoir basin, including upstream and downstream sections, main and tributary basins, and reservoir cross-sections, is collected. Preprocessing steps such as time-series alignment, missing value repair, and outlier removal are completed in a unified manner. Continuous time-series water level data from each monitoring station are accurately extracted. Using the station as the spatial dimension and the sampling time as the temporal dimension, a multi-dimensional hydrological baseline sequence covering the entire basin is constructed as the original base data for subsequent spatiotemporal feature mining.

[0054] The formula for constructing the multidimensional hydrological basic sequence matrix is: (1) In formula (1), This is a multidimensional hydrological baseline sequence matrix, with units in meters (m), sourced from the reservoir basin. Real-time water level data collected from several hydrological monitoring stations, with dimensions of [missing information]. This represents the baseline for the temporal variation of water levels at multiple stations across the entire basin; For the first monitoring stations The water level value at any given time, in meters (m), is obtained by real-time sampling from a water level sensor, and the value range is [range missing]. m; This refers to the monitoring station serial number, which has no unit and a range of values. This is a unique identifier for hydrological monitoring stations within the basin; This represents the total number of monitoring stations in the watershed, without units, and the range of values ​​is [value range missing]. In this embodiment, 16 is used; To monitor time series moments, the unit is hours (h), the sampling period is 1 hour (h), and the value range is... ; This represents the total duration of time-series sampling, in hours (h), with a value range of [value range missing]. h; `<matrix>` is a matrix transpose operator with no unit, used to transpose matrices of station and time dimensions to form a standard time-series matrix format. The control logic of formula (1) is to combine the real-time water level data of each station in row vector form according to the monitoring station number and sampling time sequence, and then transpose the result to construct a matrix with dimension `<matrix>`. The multidimensional hydrological baseline sequence matrix, with a unified data format, provides standardized input for the subsequent extraction of temporal fluctuation characteristics and spatial confluence characteristics. Formula (1) adopts a multi-site matrix data organization method, which is different from the traditional single-site time series analysis method. At the same time, it retains the spatial differences (water level distribution at different sites) and temporal continuity (water level changes at the same site) of the basin water level data, providing a complete data foundation for the subsequent decoupling of spatiotemporal dual-dimensional features, and avoiding the one-sidedness of single-site data and the problem of missing spatial information.

[0055] Multi-site hydrological monitoring data is a time-series hydrological monitoring data covering the entire area of ​​the reservoir, including upstream tributaries, reservoir area, downstream sections, and confluence nodes. The core data includes hourly water level, inflow, outflow, rainfall intensity, and watershed confluence velocity. The parameter range is: water level 0m to the reservoir's design maximum flood level, inflow 0m³ / s to 2000m³ / s, and hourly rainfall 0mm to 100mm. The data is based on the fact that this range covers hydrological changes under all conditions during normal water periods, high water periods, and rainstorm flood periods, and is compatible with hydrological monitoring standards for small, medium, and large reservoir watersheds.

[0056] Threshold for the number of monitoring stations: ≥6 monitoring stations in the watershed, preferably 6 to 12; Basis for setting the value: fewer than 6 stations cannot cover the upstream and downstream of the watershed and the confluence space structure of the main and tributary gullies, and the spatial confluence feature extraction will be distorted; 12 stations or less can take into account both the integrity of spatial representation and the efficiency of data processing, which is in line with the hydrological monitoring deployment specifications for reservoir watersheds.

[0057] The multidimensional hydrological baseline sequence is a two-dimensional water level time series constructed with watershed monitoring stations as the spatial dimension and fixed sampling intervals as the time dimension. It is a structured raw hydrological data sequence. The sequence dimension is [number of stations, time series step size], and the time series step size ranges from 24 to 72 steps, with 48 steps being the preferred size.

[0058] Data sampling timing parameters: sampling interval 5min to 30min, preferably 10min; the basis for the value is that a 10min sampling interval can accurately capture transient water conditions such as short-term rainstorms and rapid confluence, and avoid the problems of losing short-term abnormal features in low-frequency sampling and redundant computing power in high-frequency sampling.

[0059] Step S120: Obtain the temporal fluctuation feature vector based on the multidimensional hydrological baseline sequence.

[0060] We perform time-series difference operations, trend fitting, and fluctuation amplitude calculation on multidimensional hydrological basic sequences to explore the fluctuation patterns, fluctuation rates, fluctuation amplitudes, and time-series cumulative change characteristics of water levels at individual stations over time. We normalize and compress the time-series characteristics of each station to generate a unified-dimensional global time-series fluctuation feature vector, which accurately represents the dynamic evolution characteristics of hydrological time dimension.

[0061] The formula for constructing the temporal fluctuation feature vector is: (2) (3) In formulas (2)~(3), For the first Site The deviation of water level rise and fall at any given time, expressed in meters (m), represents the magnitude of the instantaneous water level deviating from the steady-state value; For the first monitoring stations The water level at any given time is expressed in meters (m), consistent with the definition in formula (1); For the first The historical steady-state average water level of the station, in meters (m), is derived from the historical hydrological statistics of the past 30 days and represents the long-term steady-state baseline of the station's water level. The station number is the monitoring station number, which has no unit and is consistent with the definition in formula (1); To monitor time series moments, the unit is hours (h), consistent with the definition in formula (1); The total number of monitoring stations in the basin is unitless and consistent with the definition in formula (1); The first vector is a time-series fluctuation feature vector with no unit. The second vector is a normalized feature output by the LSTM network with a dimension of 64, representing the time-series fluctuation pattern of water level across the entire basin. The function is a long short-term memory time-series feature extraction function with no unit. It is used to capture long-term time-dependent fluctuation features. The input is the water level fluctuation deviation sequence of each station. The control logic of formulas (2) to (3) is to first remove the steady-state water level baseline of each station through formula (2) to obtain the water level deviation sequence containing only dynamic fluctuation components; then input the deviation sequence into the LSTM network, capture the long-term dependence of the water level sequence through the gating mechanism, extract the refined time-series fluctuation hidden features, and form the time-series fluctuation feature vector. Formulas (2) to (3) eliminate the static topographic water level difference between different stations through baseline difference preprocessing, retain only the fluctuation components that reflect the dynamic changes in hydrology, enhance the identification of dynamic fluctuation features, effectively avoid the interference of static baseline on time-series feature extraction, and significantly improve the effectiveness of time-series fluctuation features and the accuracy of subsequent anomaly identification.

[0062] The temporal fluctuation feature vector is a one-dimensional feature vector that quantitatively represents the temporal dynamic changes of the water situation in the entire region. It includes quantitative information such as hourly water level rise, temporal fluctuation variance, fluctuation period, and cumulative water level deviation. The vector dimension is uniformly set to 128 dimensions, and the feature normalization interval is [0, 1].

[0063] Step S130: Use a graph convolutional network to process the spatial topology and temporal fluctuation feature vectors to obtain the spatial confluence feature matrix.

[0064] A spatial topological adjacency matrix of watershed monitoring stations is constructed to represent the spatial correlation and confluence and conduction relationships between upstream and downstream areas and between main and tributary watersheds. The spatial topological matrix of stations and the temporal fluctuation feature vectors are input into a graph convolutional network (GCN) to mine the spatial correlation characteristics of the watershed, the cross-station confluence and conduction laws, and the spatial hydrological linkage characteristics. The output is a two-dimensional structured spatial confluence feature matrix, realizing a high-dimensional mapping from temporal features to spatial topological features.

[0065] The formula for calculating the spatial confluence characteristic matrix is: (4) In formula (4), This is a spatial confluence feature matrix, without units, with dimensions of . It characterizes the spatial correlation features of confluence, runoff, and runoff among stations; This is a graph convolution spatial feature extraction function with no unit, used to mine spatial flow associations under irregular watershed topology; This is a topological adjacency matrix of watershed stations, unitless, constructed manually, representing the spatial relationships between upstream and downstream stations, with dimensions of [missing information]. ; This is the temporal fluctuation feature vector, which is unitless and consistent with the definition in formula (3), with a dimension of . ; The total number of monitoring stations in the basin is unitless and consistent with the definition in formula (1); The dimension is a single-site feature dimension, without units, and the initial dimension is 128. The control logic of formula (4) is based on the real watershed topology (derived from the watershed site topological adjacency matrix). The temporal fluctuation feature vector is used as the node feature input to the graph convolutional network. The node features and topological relationships are fused through graph convolution operations to model the spatial confluence coupling relationship between multiple stations and output the spatial confluence feature matrix. Formula (4) introduces a graph convolutional network to adapt to the irregular spatial topology of the watershed, which overcomes the defect that the traditional Euclidean spatial convolution cannot adapt to the discrete distribution of hydrological stations. It can accurately capture the confluence, runoff and runoff correlation features between upstream and downstream stations in the watershed, and greatly improve the accuracy and effectiveness of spatial confluence feature extraction.

[0066] The Graph Convolutional Network (GCN) adopts a hydrological spatially adapted two-layer graph convolutional architecture: 1. Topology building layer: generates an adjacency weight matrix of watershed stations to quantify the spatial confluence correlation strength; 2. Feature mapping layer: maps temporal fluctuation features to the spatial topology dimension; 3. Convolutional aggregation layer: aggregates hydrological correlation features of neighboring stations to extract spatial confluence patterns; 4. Matrix output layer: generates a global spatial confluence feature matrix.

[0067] Spatial topology is a spatial network structure composed of the geographical location associations, upstream and downstream confluence sequences, and regional connectivity of various monitoring stations in a reservoir basin. It is used to quantify the spatial transmission relationships of hydrology.

[0068] Spatial confluence feature matrix: A two-dimensional feature matrix in which rows correspond to spatial monitoring stations and columns correspond to spatial confluence feature dimensions. It is used to characterize the confluence distribution, conduction intensity, and spatial anomaly distribution characteristics of the entire watershed.

[0069] Step S140: Determine whether the dimension of the spatial confluence feature matrix is ​​greater than the preset dimension threshold.

[0070] The feature dimensions of the spatial confluence feature matrix output by step S130 are statistically analyzed and compared with a preset dimension threshold to distinguish between high-dimensional redundant matrices and low-dimensional effective matrices, providing a basis for adaptive dimensionality reduction processing and balancing feature integrity and computational efficiency.

[0071] The formula for determining the feature dimension threshold is: (5) In formula (5), This is a matrix dimension calculation function, which is dimensionless and used to calculate the eigendimensionality of a matrix. This is the spatial confluence characteristic matrix, which is unitless and consistent with the definition of formula (4); The preset dimension threshold is unitless; in this embodiment, it is set to 256, used to determine whether feature redundancy exists. The control logic of formula (5) is to determine whether the dimension of the spatial confluence feature matrix is ​​greater than the preset dimension threshold. If it is greater, feature redundancy is determined, and dimension reduction processing is triggered; if it is not greater, the spatial confluence feature matrix is ​​directly output to balance the amount of feature information and subsequent computational efficiency. Formula (5) achieves adaptive redundancy filtering of spatial confluence features through dimension threshold determination. While retaining key hydrological correlation information, it effectively reduces the feature dimension, avoids computational redundancy and overfitting risks caused by high-dimensional features, and improves the efficiency and stability of subsequent spatiotemporal feature decoupling and anomaly identification.

[0072] The preset dimension threshold is the threshold for judging the feature dimension of the spatial confluence feature matrix, with a value of 256. Judgment rules: if the feature dimension of the spatial confluence feature matrix is ​​greater than 256, it is judged as a high-dimensional redundant matrix and triggers dimensionality reduction processing; if the feature dimension of the spatial confluence feature matrix is ​​less than or equal to 256, it is judged as dimensionally compliant and can be directly used as a valid spatial feature.

[0073] Value selection criteria: The effective information dimension of hydrological spatial features is concentrated within 256 dimensions. Any new features beyond this dimension are redundant information and noise, which will increase the computational load of subsequent models and reduce decoupling accuracy. 256 dimensions is the optimal critical threshold that balances feature integrity and computational efficiency.

[0074] Step S150: If the dimension of the spatial confluence feature matrix is ​​greater than the preset dimension threshold, then the spatial confluence feature matrix is ​​reduced in dimension to obtain the target spatial confluence feature.

[0075] When the spatial confluence feature matrix exceeds the dimension limit and has dimensional redundancy, principal component analysis (PCA) is used to adaptively reduce the dimensionality of the high-dimensional matrix, retaining the core spatial confluence features with the top 99% contribution rate, eliminating redundant noise dimensions, and refining the target spatial confluence features to obtain low-dimensional, highly recognizable features.

[0076] The formula for calculating the target space confluence characteristics is: (6) In formula (6), The target spatial convergence feature is unitless and has a dimension of 64 after dimensionality reduction, making it the optimal feature for refined spatial convergence. Principal component analysis dimensionality reduction function, unitless, retains more than 95% of effective feature information; This is the spatial confluence characteristic matrix, which is unitless and consistent with the definition of formula (4); The preset dimension threshold is unitless and consistent with the definition in formula (5). In this embodiment, the value is 256, which is taken as the target dimension after dimensionality reduction. The control logic of formula (6) is that when the dimension of the spatial confluence feature matrix is ​​greater than the preset dimension threshold, principal component analysis (PCA) is used to reduce and compress the high-dimensional redundant feature matrix. As the target dimension, more than 95% of the effective feature information is retained, and noise and redundant information are removed to obtain low-dimensional, high-information-density target spatial confluence features. Formula (6) achieves efficient compression of spatial confluence features through adaptive dimension filtering and principal component analysis dimensionality reduction. While preserving the integrity of core hydrological correlation information, it significantly reduces the feature dimension, taking into account both the spatial feature expression capability and the computational efficiency of subsequent spatiotemporal decoupling algorithms, and effectively avoiding the overfitting risk brought about by high-dimensional features.

[0077] Dimensionality reduction is a structured computational process that removes redundancy and purifies core features from a high-dimensional spatial feature matrix, preserving effective hydrological spatial correlation features and eliminating dimensional redundancy interference.

[0078] The target space convergence feature is a standardized space convergence feature after dimensionality reduction and purification, with a unified dimension of 256 dimensions, a feature information retention rate of ≥99%, and a noise residual rate of ≤1%.

[0079] Step S160: Use a long short-term memory network to process the temporal fluctuation feature vector and the target space confluence feature to obtain the initial hydrological feature set.

[0080] The purified temporal fluctuation feature vectors are temporally aligned with the target spatial confluence features and input into an improved long short-term memory network to complete deep fusion of spatiotemporal features, secondary stripping of redundancy, and feature association enhancement. Finally, the system integrates and generates a complete, loosely coupled, and highly accurate initial hydrological feature set with both temporal and spatial dimensions.

[0081] The formula for calculating the initial hydrological characteristic set is: (7) In formula (7), The initial set of hydrological features is unitless, while the global set of hydrological features, which integrates temporal fluctuations and spatial confluence, has a dimension of 128. This is a feature splicing and fusion function with no unit, used to achieve dimensional alignment and channel-level fusion of temporal and spatial features; It is the time series fluctuation feature vector, which has no unit and is consistent with the definition of formula (3), with a dimension of 64; The target spatial confluence feature is unitless and consistent with the definition in formula (6), with a dimension of 64. The control logic of formula (7) is to fuse the temporal fluctuation feature vector with a dimension of 64 and the target spatial confluence feature with a dimension of 64 through a feature concatenation function to obtain an initial hydrological feature set with a dimension of 128, unifying the temporal and spatial feature dimensions and constructing a complete initial hydrological spatiotemporal feature system. Formula (7) realizes the integrated fusion of the temporal fluctuation feature vector and the target spatial confluence feature through feature concatenation, forming a global hydrological feature that simultaneously includes the dynamic changes in the time dimension and the correlation in the spatial dimension. This provides complete and unified input data for subsequent spatiotemporal feature decoupling processing, avoiding the information fragmentation problem caused by processing the two types of features separately.

[0082] Complete architecture of Long Short-Term Memory (LSTM) network: 1. Dual Feature Alignment Layer: Unifies the timestamp and dimension standards of temporal and spatial features; 2. Feature Fusion Encoding Layer: Remembers the long-term correlation patterns of spatiotemporal features; 3. Redundancy Refinement Layer: Removes residual coupling interference; 4. Feature Set Output Layer: Outputs a standardized initial hydrological feature set.

[0083] The initial hydrological feature set is a two-dimensional basic feature dataset composed of temporal fluctuation features and target spatial confluence features in a structured way. It has unified feature dimensions, complete temporal and spatial information, and noise ratio ≤1%, providing high-quality feature support for subsequent accurate decoupling.

[0084] Preferably, the reservoir hydrological spatiotemporal feature decoupling and hierarchical early warning method provided in this embodiment includes step S200 as follows: Step S210: The temporal fluctuation features and spatial confluence features in the initial hydrological feature set are spliced ​​together to obtain a spatiotemporal mixed signal matrix.

[0085] The temporal fluctuation feature vectors in the initial hydrological feature set are dimensionally aligned and horizontally stitched with the target spatial confluence features. The temporal and spatial dimensional information is integrated to construct a spatiotemporal hybrid signal matrix containing temporal-spatial coupling correlation information. The original coupling feature information is fully preserved, providing the original input matrix for subsequent decoupling operations.

[0086] The formula for constructing a spatiotemporal mixed signal matrix is: (8) In formula (8), It is a spatiotemporal mixed signal matrix, without dimensions, with dimensions of . It represents a hybrid signal that combines temporal and spatial characteristics; This is the temporal fluctuation feature vector, which is unitless and consistent with the definition in formula (3), with a dimension of . ; The target space convergence feature is unitless, consistent with the definition of formula (6), and has a dimension of . ; `x` is a matrix transpose operator with no unit, used to convert row vectors into column vectors to achieve dimension alignment. The control logic of formula (8) is to convert the row vector into a column vector with dimension `x`. The temporal fluctuation feature vector and dimension are The target space confluence feature vectors are transposed into column vectors, and then concatenated column by column to form the dimension. The matrix is ​​used to construct a standard mixed observation signal matrix, which is adapted to the input specifications of the Independent Component Analysis (ICA) algorithm. Formula (8) transforms the complex hydrological spatiotemporal coupling problem into a standard blind source separation problem by constructing a standardized coupled signal matrix, providing a standardized input form for subsequent spatiotemporal feature decoupling using the Independent Component Analysis algorithm, and solving the problem of difficult separation of spatiotemporal feature coupling in traditional methods.

[0087] The spatiotemporal hybrid signal matrix is ​​a two-dimensional coupling matrix that integrates temporal fluctuation characteristics and spatial confluence characteristics. It carries all the original information of hydrological temporal fluctuations, spatial confluence transmission, and multi-dimensional coupling interference, and is the basic operation matrix for feature decoupling.

[0088] Step S220: Perform an orthogonal transformation on the spatiotemporal mixed signal matrix to obtain the whitening feature matrix.

[0089] The spatiotemporal mixed signal matrix is ​​decentered and orthogonally whitened to eliminate matrix feature correlation, unify the variance distribution of features in each dimension, and weaken dimensional coupling interference. This transforms the coupled mixed matrix into a standardized whitened feature matrix, improving the decoupling accuracy and convergence speed of subsequent independent component analysis.

[0090] The formula for calculating the whitening feature matrix is: (9) (10) In formulas (9)~(10), The covariance matrix of the spatiotemporal mixed signal is dimensionless and represents the coupling correlation between temporal fluctuation characteristics and spatial confluence characteristics. Its dimension is [missing information]. ; This is a mathematical expectation operation, without units, used to calculate the statistical average of matrix multiplication; This is a spatiotemporal mixed signal matrix, without units, consistent with the definition in formula (8), and with dimensions of... ; The first matrix is ​​a diagonal matrix of covariance eigenvalues, without units, and the second matrix is ​​the covariance matrix of the spatiotemporal mixed signal. A diagonal matrix composed of the eigenvalues; The eigenvector matrix is ​​a unitless covariance matrix, representing the covariance matrix of the spatiotemporal mixed signal. The orthogonal eigenvector matrix; The whitening feature matrix is ​​dimensionless, and the standard feature matrix is ​​obtained after orthogonal transformation, which is orthogonally independent and variance-normalized, with dimensions of . The control logic of formulas (9) to (10) is to first calculate the spatiotemporal mixed signal covariance matrix of the spatiotemporal mixed signal matrix using formula (9) to quantify the coupling correlation between temporal and spatial features; then, to perform eigenvalue decomposition on the spatiotemporal mixed signal covariance matrix to obtain the diagonal matrix of covariance eigenvalues. With covariance eigenvector matrix Finally, the spatiotemporal mixed signal matrix is ​​orthogonally whitened using formula (10) to eliminate the correlation between features and normalize the variance, thus obtaining a whitened feature matrix that satisfies the convergence condition of the ICA algorithm. Formulas (9) to (10) address the problem of strong coupling and high correlation of hydrological spatiotemporal features through orthogonal whitening, providing an orthogonal input basis for subsequent independent component analysis algorithms, effectively improving the accuracy and stability of spatiotemporal feature decoupling, and avoiding the convergence difficulties and decoupling errors of the ICA algorithm caused by feature coupling.

[0091] Orthogonal transformation is a standardized preprocessing operation for coupled matrices, achieving orthogonal independence of features, dimensional decoupling preprocessing, and variance normalization.

[0092] The whitening feature matrix is ​​a standardized matrix after orthogonal transformation. The features in each dimension are orthogonal to each other and have no linear correlation. The matrix variance is uniformly 1, the mean is 0, and the decoupling preprocessing accuracy is ≥99%.

[0093] Step S230: If the whitening feature matrix satisfies the preset full rank condition, then the independent component analysis algorithm is used to process the whitening feature matrix to obtain the unmixed matrix.

[0094] Verify the rank property of the whitening feature matrix to determine if it meets the preset full rank condition; under the premise that the matrix is ​​full rank and the feature information is complete and without missing parts, start the independent component analysis iterative operation to solve the optimal solution mixing matrix that can separate temporal and spatial coupled components, and realize the blind source separation mapping of mixed features.

[0095] The formulas for determining full rank and solving for the unmixing matrix are as follows: (11) In formula (11), This is a matrix rank calculation function, which is unitless and used to calculate the rank of the whitening characteristic matrix. The whitening feature matrix is ​​unitless and consistent with the definition in formula (10), with dimensions of . ; is the effective dimension of the feature, has no unit, takes a value of 2, and is the number of rows in the whitening feature matrix; The unmixing matrix is ​​dimensionless, and the core decoupling weight matrix is ​​obtained through iterative solving using the independent component analysis algorithm. Its dimension is [missing information]. ; This is the unmixing operation function for Independent Component Analysis (ICA), which is unitless and used to solve for the unmixing matrix of independent components from the whitened feature matrix. The control logic of formula (11) is to first determine whether the rank of the whitened feature matrix is ​​equal to the effective dimension of the feature by using the matrix rank solving function to verify whether the whitened feature matrix is ​​full rank; if the full rank condition is met, the whitened feature matrix is ​​determined to be valid, and the ICA algorithm is input to iteratively solve for the optimal unmixing matrix, providing core weight parameters for subsequent spatiotemporal feature decoupling. Formula (11) introduces a full rank verification mechanism to determine the validity of the input whitened feature matrix before ICA unmixing, avoiding the decoupling failure problem caused by singular matrices or reduced-rank matrices, greatly improving the stability and robustness of the spatiotemporal feature decoupling algorithm, and solving the problem of insufficient matrix rank caused by noise or anomalies in hydrological data.

[0096] Preset full-rank condition: The row rank of the whitened feature matrix is ​​equal to the column rank, and the rank value is ≥ 98% of the matrix dimension; Value basis: A full-rank matrix represents complete feature information, no dimension collapse, and no information loss, which can guarantee that independent components are decoupled without distortion; Non-full-rank matrices have feature redundancy and missing features, the decoupling result is invalid, and data needs to be collected and processed again.

[0097] The demixing matrix is ​​the optimal transformation matrix obtained through iterative solving of the ICA algorithm. It is used to achieve linear decoupling and separation of spatiotemporally mixed signals. The matrix iteration convergence threshold is 10. -6 ~10 -4 10 preferred -5 The maximum number of iterations is 80 to 150, with 100 being the optimal number.

[0098] Step S240: Perform a linear transformation on the spatiotemporal mixed signal matrix using the demixing matrix to obtain independent component components, and generate spatiotemporal decoupling difference features based on the independent component components.

[0099] Linear transformation is performed on the spatiotemporal mixed signal matrix and the demixing matrix to separate the pure temporal independent component and the pure spatial independent component. The deviation between the decoupled pure features and the historical steady-state benchmark features is calculated to quantify the temporal fluctuation deviation and the spatial convergence anomaly deviation, and then integrated to generate spatiotemporal decoupling difference features.

[0100] The formula for generating spatiotemporal decoupling difference features is: (12) (13) In formulas (12)~(13), It is an independent component matrix, without dimensions, with dimensions of . Includes time-independent components Spatial independent components ; The unmixing matrix is ​​unitless, consistent with the definition in formula (11), and has dimensions of . ; This is a spatiotemporal mixed signal matrix, without units, consistent with the definition in formula (8), and with dimensions of... ; It is a time-independent component, without units, and is a pure time-fluctuation characteristic component separated from the mixed signal; It is a spatially independent component with no unit, representing a pure spatial confluence characteristic component separated from a mixed signal; The difference between the temporal and spatial independent components is a unitless feature representing the absolute difference between the temporal and spatial independent components. It characterizes the degree of independence after the temporal and spatial features are decoupled and is the core output feature of this step. The control logic of formulas (12) and (13) is to first use formula (12) to perform a linear transformation on the temporal and spatial mixed signal matrix using the demixing matrix to achieve blind source separation of the coupled signal and obtain mutually independent temporal and spatial component components; then, the temporal independent components are calculated using formula (13). Spatial independent components The absolute difference between the two independent components generates spatiotemporal decoupling difference features, thus completing the spatiotemporal feature decoupling processing of hydrological data. Formulas (12) to (13) are the first to apply ICA blind source separation technology to the spatiotemporal feature decoupling of reservoir hydrological data, completely removing the coupling interference between temporal fluctuation features and spatial confluence features, and obtaining pure and independent spatiotemporal components and their difference features. This provides a highly reliable input basis for subsequent hydrological anomaly identification and graded early warning, and solves the problem of low anomaly identification accuracy caused by spatiotemporal feature coupling in traditional methods.

[0101] Independent components are pure temporal and spatial feature components after complete removal of coupling interference. The two types of components are independent of each other and have no coupling or aliasing, with a decoupling accuracy of ≥98.5%.

[0102] The spatiotemporal decoupling difference feature is an integrated set of temporal fluctuation deviation features and spatial confluence deviation features. It is used to quantify the temporal fluctuation deviation and spatial confluence anomaly deviation of the current hydrological state relative to the historical steady-state hydrological state. The feature deviation range is [0, 1], and the larger the value, the higher the degree of hydrological anomaly.

[0103] Furthermore, the reservoir hydrological spatiotemporal characteristic decoupling and hierarchical early warning method provided in this embodiment includes step S300 as follows: Step S310: Obtain the spatiotemporal decoupling difference features and calculate the spatiotemporal decoupling feature fluctuation amplitude of the spatiotemporal decoupling difference features.

[0104] Read the standardized spatiotemporal decoupling difference features output in step S200, and through time series variance calculation and spatial deviation gradient calculation, comprehensively solve to obtain the spatiotemporal decoupling feature fluctuation amplitude that can uniformly characterize the severity of time series fluctuations and the degree of spatial anomaly deviation, as the core quantitative indicator for anomaly judgment.

[0105] The formula for calculating the amplitude of spatiotemporal decoupling characteristic fluctuations is: (14) In formula (14), The amplitude of the spatiotemporal decoupling characteristic fluctuation is dimensionless and represents the severity of abnormal fluctuations in water conditions. The number of feature sampling points is unitless, and the total number of dimensions of the spatiotemporally decoupled differential features is also represented. For the first The dimensionless decoupling difference eigenvalue is the th element of the spatiotemporal decoupling difference eigenvalue matrix. The elements, as defined in formula (13) correspond; The mean of the decoupling features is dimensionless, representing the arithmetic mean of all elements in the spatiotemporal decoupling difference feature matrix, i.e.: The control logic of formula (14) is to first calculate the mean of the decoupling features of the spatiotemporal decoupling difference feature matrix. Then, sum the absolute deviations of the decoupling difference eigenvalues ​​and the mean of the decoupling features for each dimension, and finally divide by the number of feature sampling points. The spatiotemporal decoupling characteristic fluctuation amplitude in the form of mean absolute deviation is obtained. Formula (14) quantifies the intensity of abnormal fluctuations in hydrological data. It uses the mean absolute deviation index to quantify the fluctuation amplitude of spatiotemporal decoupling difference characteristics, avoiding the shortcomings of traditional variance-based calculations that are sensitive to outliers. It can more robustly reflect the intensity of abnormal fluctuations in hydrological data and provides a reliable quantitative indicator for subsequent hydrological anomaly classification and early warning.

[0106] The spatiotemporal decoupling characteristic fluctuation amplitude is a comprehensive quantitative index that integrates the temporal fluctuation variance and the spatial deviation gradient, with a value range of [0, 1], used to uniformly evaluate the degree of spatiotemporal comprehensive abnormal fluctuations in water conditions.

[0107] Step S320: If the fluctuation amplitude of the spatiotemporal decoupling feature exceeds the preset feature fluctuation threshold, the spatiotemporal decoupling difference feature is segmented to obtain the evolution trend sequence.

[0108] The fluctuation amplitude of the spatiotemporal decoupling feature is compared with the preset feature fluctuation threshold. If it is determined to exceed the limit, it is confirmed that there is an abnormal nascent trend in the current water situation. The continuous spatiotemporal decoupling difference features are segmented according to a fixed time-series sliding window to generate multiple continuous time-series evolution trend sequences for subsequent model trend inference.

[0109] The formula for segmenting abnormal situation sequences is: (15) In formula (15), The amplitude of the spatiotemporal decoupling characteristic fluctuation is unitless and consistent with the definition of formula (14); The preset characteristic fluctuation threshold has no unit; in this embodiment, it is set to 0.35, which is used to determine whether there are abnormal fluctuations in the hydrological data. It is an evolutionary trend sequence, without units, and is a set of time series subsequences of anomalous features, composed of multiple segmented subsequences; This is a sequence segmentation function with no unit, used to segment a high-dimensional feature vector into multiple continuous subsequences of fixed length; The difference between spatiotemporal decoupling is a unitless feature, consistent with the definition of formula (13); The length of a single subsequence is unitless; in this embodiment, it is set to 12, representing the number of feature points contained in each subsequence. The control logic of formula (15) is to first determine whether the fluctuation amplitude of the spatiotemporal decoupling feature exceeds the preset feature fluctuation threshold. If it does, it is determined that there is abnormal fluctuation in the current hydrological data. Then, the spatiotemporal decoupling difference features are divided into segments of fixed length using a sequence segmentation function. The sequence is divided into multiple continuous subsequences to generate an evolution trend sequence, which is adapted to the input specifications of the subsequent time series prediction model. Formula (15) achieves accurate identification and standardized input of water situation anomalies through threshold judgment and sequence segmentation, converts continuous abnormal features into fixed-length time series subsequences, and provides a standardized and directly input data format for subsequent water situation evolution trend prediction, solving the problem that abnormal features in traditional water situation early warning cannot be directly adapted to the time series model.

[0110] Preset feature fluctuation threshold: The safe threshold for the fluctuation amplitude of spatiotemporal decoupling features is 0.35; Judgment rules: If the fluctuation amplitude is >0.35, it is judged as feature fluctuation exceeding the limit, indicating a risk of abnormal water conditions, and the situation identification process is initiated; If the fluctuation amplitude is ≤0.35, it is judged as water conditions in a steady state and normal.

[0111] The preset characteristic fluctuation threshold is determined based on the statistical analysis of the reservoir's historical hydrological steady-state data over many years. During normal water periods and regular high water periods, the fluctuation range of hydrological characteristics is stable within the range of 0 to 0.35, which is within the range of natural and normal hydrological fluctuations. When the deviation exceeds 0.35, it indicates the presence of unconventional disturbances such as sudden surges in water levels due to short-term heavy rainfall, concentrated upstream flow, and abnormal rises in water levels, which have the potential to evolve into dangerous situations. If the preset characteristic fluctuation threshold is too high, it will miss the initial signs of weak anomalies, while if the preset characteristic fluctuation threshold is too low, it will cause high-frequency false alarms. 0.35 is the optimal critical threshold that balances sensitivity and accuracy.

[0112] The evolution trend sequence is a continuous subsequence of abnormal features after time-series sliding window segmentation. The preferred duration of a single window is 10 minutes, and the step size of a single sequence is 30 steps. It is used to characterize the continuous evolution process of abnormal water conditions.

[0113] Step S330: Input the evolution trend sequence into the pre-constructed spatiotemporal dual-dimensional evaluation model for feature extraction and generate fluctuation evolution vectors.

[0114] The standardized evolution sequence is input into the training convergent spatiotemporal dual-dimensional evaluation model. The abnormal temporal fluctuation patterns and spatial confluence and diffusion patterns are extracted synchronously through the model's temporal coding layer and spatial convolutional layer. The high-dimensional fluctuation evolution vector is compressed and generated to centrally carry the dynamic evolution characteristics of water anomalies.

[0115] The formula for generating fluctuation evolution vectors is: (16) In formula (16), is the fluctuation evolution vector, which has no unit; is the temporal evolution feature output by the spatiotemporal dual-dimensional evaluation model, with a dimension of 32. It is a pre-built spatiotemporal dual-dimensional evaluation model without units. It is a dedicated model for extracting hydrological anomalies and has built-in temporal convolution and self-attention modules. The evolutionary trend sequence is unitless and consistent with the definition in formula (15), representing the set of time-series subsequences of anomalous features after segmentation. The control logic of formula (16) is to input the evolutionary trend sequence into a pre-trained spatiotemporal dual-dimensional evaluation model, extract local anomalous patterns through the model's temporal convolutional layer, capture long-term temporal dependencies through the self-attention module, and finally output a fluctuation evolution vector with a dimension of 32, thus completing the characteristic expression of the hydrological anomalous trend. Formula (16) achieves in-depth mining of the time-series features of hydrological anomalies through a dedicated spatiotemporal dual-dimensional evaluation model. It captures both local fluctuation patterns and models long-term temporal dependencies. The generated fluctuation evolution vector can comprehensively characterize the evolution trend of hydrological anomalies, providing a highly discriminative feature basis for subsequent graded early warning.

[0116] The fluctuation evolution vector is the core feature vector of abnormal situation extracted by the spatiotemporal dual-dimensional assessment model. It quantitatively represents the abnormal fluctuation rate, evolution trend, and spatial diffusion intensity of water conditions, and the dimension is unified to 64 dimensions.

[0117] Step S340: Map the fluctuation evolution vector to obtain the situation fusion features.

[0118] A high-dimensional nonlinear mapping transformation is performed on the fluctuation evolution vector to achieve deep fusion of temporal and spatial anomaly features, redundant secondary purification, elimination of the one-sidedness of single-dimensional features, and generation of global situational fusion features that take into account both temporal and spatial dimensions.

[0119] The situation fusion feature mapping formula is: (17) In formula (17), It is a situational fusion feature, without units, and is a unified anomaly feature fused from spatiotemporal dimensions; It is a non-linear activation function with no unit. In this embodiment, the Sigmoid function is used to introduce a non-linear transformation. Here, the situation mapping weight matrix is ​​dimensionless, and the weights are pre-trained linear transformation weights with dimension 1. ; It is the fluctuation evolution vector, which has no unit and is consistent with the definition of formula (16), with a dimension of 32; Let be the situation bias vector, which is unitless, and be the pre-trained bias parameters, with dimension . The control logic of formula (17) is to first map the situation to the weight matrix. and situation bias vector A linear transformation is applied to the fluctuation evolution vector, followed by a nonlinear activation function. By introducing nonlinear mapping, high-dimensional temporal evolution features are transformed into unified-dimensional situational fusion features, thus realizing the final aggregation of spatiotemporal anomaly information. Formula (17) transforms high-dimensional temporal evolution features into unified situational fusion features through a combination of linear transformation and nonlinear activation. This not only eliminates the distribution differences of features in different dimensions, but also introduces nonlinear expressive power, enhancing the distinguishability of anomaly features and providing standardized and robust feature inputs for subsequent hydrological anomaly classification.

[0120] Situation fusion features are the core features of the whole domain after the fusion of spatiotemporal dual-dimensional anomaly features. They fully cover the temporal evolution and spatial distribution patterns of water situation anomalies, providing the core basis for final situation identification.

[0121] Step S350: Use the spatiotemporal dual-dimensional evaluation model to identify abnormal evolution of the situation based on the situation fusion features, and obtain the results of abnormal evolution of the situation.

[0122] Through the situation assessment and scoring output layer of the spatiotemporal dual-dimensional evaluation model, the situation fusion characteristics are quantitatively evaluated and the trend is extrapolated to predict the evolution speed, spread range and risk intensity of the water situation anomaly in the next 1 to 6 hours, and output the anomaly evolution situation result that combines quantitative scoring and situation labeling.

[0123] The formula for identifying abnormal evolution trends is: (18) In formula (18), The result represents the abnormal evolution trend, without units, and is a probability distribution including three types of trends: slow evolution, sudden evolution, and continuous deterioration. Softmax is a unitless normalization function used to convert feature vectors into probability distributions, ensuring that the sum of the probabilities of each situation is 1. The situation fusion feature is unitless and consistent with the definition in formula (17). The control logic of formula (18) is to input the situation fusion feature into the Softmax function for normalization and output the probability distribution of various abnormal evolution situations to complete the identification and classification of abnormal evolution situations of water conditions. Formula (18) performs situation identification based on the pure abnormal features after spatiotemporal decoupling, avoiding the interference of spatiotemporal coupling noise on the identification results in traditional methods, greatly improving the accuracy and reliability of the identification of abnormal evolution situations of water conditions, and providing accurate decision-making basis for subsequent graded early warning.

[0124] Complete architecture of the spatiotemporal dual-dimensional evaluation model: The system adopts a five-layer architecture: 1. Sequence Input Layer: imports the evolution trend sequence; 2. Spatiotemporal Feature Extraction Layer: extracts temporal and spatial anomaly features in parallel; 3. Vector Mapping Layer: generates fluctuation evolution vectors; 4. Feature Fusion Layer: outputs trend fusion features; 5. Trend Quantization Output Layer: outputs the results of anomaly evolution trends.

[0125] The abnormal evolution situation results are structured results that quantitatively characterize the severity level, evolution rate, spatial diffusion range, and duration of the water situation anomaly. They include four types of situation labels: slow evolution, rapid evolution, local confluence anomaly, and global confluence anomaly, with a score range of [0, 100]. The higher the score, the greater the risk of the danger.

[0126] Prediction time range: Water situation forecast for the next 1 to 6 hours, with a preferred forecast window of 3 hours; Basis for selection: 3 hours is the golden forecast time for reservoir flood control scheduling, which can meet the engineering requirements of early warning and reserved scheduling buffer time.

[0127] Preferably, the reservoir hydrological spatiotemporal feature decoupling and hierarchical early warning method provided in this embodiment includes step S400 as follows: Step S410: Obtain the current reservoir capacity status data of the target reservoir, normalize the current reservoir capacity status data, and generate a standard reservoir capacity status matrix.

[0128] Real-time data collection of core reservoir status data, such as real-time reservoir capacity, reservoir capacity fullness rate, remaining flood control capacity, and difference between flood control and flood control levels, eliminates dimensional differences, completes normalization and standardization processing of the [0,1] interval, and constructs a standard reservoir status matrix with unified dimensions and regular data, providing standardized input for subsequent matching operations.

[0129] The formula for constructing the standard warehouse capacity state matrix is: (19) (20) In formulas (19)~(20), This is the original storage capacity state vector, in m³, a three-dimensional column vector containing the current real-time storage capacity. Rated reservoir capacity Remaining flood control adjustable reservoir capacity ; The current real-time reservoir capacity is expressed in m³, representing the actual amount of water currently stored in the reservoir. The rated capacity of the reservoir is expressed in m³, while the maximum allowable water storage capacity of the reservoir is designed to be 100%. The remaining adjustable flood control capacity is expressed in m³, representing the difference between the rated capacity and the current real-time capacity, and characterizing the flood control adjustment margin of the reservoir. The standard storage capacity state matrix is ​​unitless and is a three-dimensional vector whose numerical range is uniformly mapped to the interval [0, 1] after min-max normalization. This is a maximum value function with no unit, used to calculate the maximum element value in the original reservoir capacity state vector; The minimum value function is unitless and is used to calculate the minimum element value in the original reservoir capacity state vector. The control logic of formulas (19) to (20) is to first construct the current real-time reservoir capacity, rated reservoir capacity, and remaining adjustable reservoir capacity data into a three-dimensional original reservoir capacity state vector, and then perform a linear transformation on each element in the vector through the min-max normalization formula to uniformly map the reservoir capacity data of different orders to the interval [0, 1], generate a standard reservoir capacity state matrix, eliminate the difference in data dimensions, and adapt to the input specifications of the subsequent graded early warning model. Formulas (19) to (20) realize the standardized representation of reservoir capacity state through the construction and normalization of the three-dimensional reservoir capacity state vector. It not only retains the relative relationship between real-time reservoir capacity, rated reservoir capacity, and adjustable reservoir capacity, but also eliminates the difference in data dimensions, providing a unified and reliable reservoir capacity state feature for subsequent graded early warning based on abnormal water conditions.

[0130] Current reservoir capacity status data is a core data set characterizing the real-time water storage capacity of a reservoir, including real-time reservoir capacity, reservoir capacity fullness rate, remaining flood control capacity, real-time water level, and difference between flood control limit water level; core parameter range: reservoir capacity fullness rate ranges from 30% to 100%, remaining flood control capacity ranges from 0 m³ to the maximum designed flood control capacity; the data is based on: covering all operating conditions of the reservoir under low water level (empty), normal water storage, and high water level (full reservoir before flood season).

[0131] The standard reservoir capacity state matrix is ​​a normalized structured reservoir capacity state matrix that uniformly represents the real-time flood control capacity and water storage status of the reservoir.

[0132] Step S420: Calculate the situation matching distance between the abnormal evolution situation result and the standard storage capacity state matrix.

[0133] The Euclidean distance algorithm is used to quantify the spatial matching distance between the abnormal evolution trend characteristics and the standard reservoir capacity characteristics, characterizing the degree of adaptation and matching between the abnormal water situation risk and the reservoir capacity carrying capacity, and realizing the quantitative correlation of the two factors of "water situation - reservoir capacity carrying capacity".

[0134] The formula for calculating the situation matching distance is: (twenty one) In formula (21), The situation matching distance is unitless and represents the degree of fit between the abnormal evolution of water conditions and the current reservoir capacity. The result of the abnormal evolution situation is unitless and consistent with the definition of formula (18), and is the probability distribution vector of various abnormal situations; The standard storage capacity state matrix is ​​unitless and consistent with the definition of formula (20). It is a normalized three-dimensional storage capacity state vector. The L2 norm distance operation is a unitless method used to calculate the Euclidean distance between two vectors. The control logic of formula (21) is to calculate the L2 norm distance between the abnormal evolution situation result vector and the standard reservoir capacity state vector to obtain the situation matching distance, thus quantifying the degree of fit between the two. The smaller the distance, the higher the fit between the current reservoir capacity state and the abnormal water situation, and the greater the risk faced by the reservoir. Formula (21) constructs a matching degree index between the situation and the reservoir capacity state through the L2 norm distance, and quantifies the information of the two different dimensions of abnormal evolution situation and reservoir capacity state in a unified manner. This provides a key indicator that can be directly used for risk assessment for subsequent graded early warning, and solves the problem of difficulty in integrating multi-dimensional information for decision-making in traditional water situation early warning.

[0135] The situation matching distance is a quantitative indicator that measures the degree of matching between abnormal water conditions and reservoir capacity. The value range is [0, 1]. The smaller the distance, the higher the coupling degree between the abnormal situation and the current reservoir capacity risk, and the greater the potential danger.

[0136] Step S430: If the situation matching distance is less than the preset matching distance threshold, extract the features of the standard reservoir capacity state matrix and construct a flood control and storage feature set.

[0137] The situation matching distance is compared with the preset matching distance threshold. When the situation matching distance is less than the preset matching distance threshold, it is determined that the water situation is highly coupled with the reservoir capacity carrying capacity and there is a risk of flood control. Then, the core storage and regulation features of the standard reservoir capacity state matrix are extracted and integrated to construct a flood control and storage feature set.

[0138] The formula for constructing the flood control and water storage feature set is: (twenty two) In formula (22), The situation matching distance is unitless and consistent with the definition in formula (21); The preset matching distance threshold has no unit; in this embodiment, it is set to 0.25, which is used to determine whether the degree of matching between the abnormal water situation and the reservoir capacity status reaches a high-risk level. The flood control and storage feature set is unitless and is a feature set extracted from the standard reservoir capacity state matrix, representing the current storage capacity of the reservoir. This is a unitless reservoir capacity feature extraction function used to extract key storage and regulation features from the standard reservoir capacity status matrix. The standard reservoir capacity state matrix is ​​unitless and consistent with the definition of formula (20). The control logic of formula (22) is to first determine whether the situation matching distance is less than the preset matching distance threshold. If it is less, it means that the current abnormal water situation and the reservoir capacity state are highly compatible, and the reservoir faces a high flood control risk. Then, the standard reservoir capacity state matrix is ​​used to extract features through the reservoir capacity feature extraction function to construct a flood control and storage feature set, which provides a basis for subsequent flood control and storage capacity assessment. Formula (22) achieves accurate identification of high-risk scenarios through the threshold judgment mechanism and constructs a flood control and storage feature set in a targeted manner. It realizes the linkage assessment of abnormal water situation and reservoir capacity state, solves the problem that traditional water situation early warning only focuses on abnormal water situation and ignores the actual storage capacity of the reservoir, and improves the scientificity and practicality of early warning decision-making.

[0139] The preset matching distance threshold is the situation matching distance judgment threshold, with a value of 0.25; the judgment rule is: if the matching distance is <0.25, it is judged as a high coupling risk state, and the reservoir capacity feature is extracted to construct the regulation set; if the matching distance is ≥0.25, it is judged as a low coupling steady state with no significant risk.

[0140] Value determination: Based on the statistical analysis of a large number of reservoir flood situation samples, 0.25 is the critical coupling threshold for the synergistic risk of abnormal water conditions and reservoir pressure. When the value is less than this, the rise and fall of water conditions will quickly exceed the reservoir's storage capacity, which may easily lead to dangerous situations. When the value is greater than this, the reservoir can autonomously buffer water condition fluctuations, and the risk is controllable.

[0141] Flood control and storage feature set: This set of core features includes remaining flood control capacity, storage capacity margin, water level safety margin, and reservoir pressure level, and is used to characterize the real-time flood control and dispatching capabilities of reservoirs.

[0142] Step S440: Classify the flood control and storage feature set to obtain the classification results for the flood season stage.

[0143] Based on the reservoir capacity fullness rate and remaining storage capacity, the flood control and storage characteristic set is classified according to the working conditions, distinguishing four stages: the stable water level period, the pre-flood preparation period, the flood season risk period, and the high water level critical period, and outputting accurate flood season stage classification results.

[0144] The formula for calculating the classification results during the flood season is as follows: (twenty three) In formula (23), The results are classified according to the flood season stage, without units, and include three categories: the initial flood season, the main flood season, and the post-flood season. This is a classification function for the flood season, without units, and represents a preset classification model or rule set. The flood control and storage feature set is unitless and consistent with the definition in formula (22). The control logic of formula (23) is to input the flood control and storage feature set into the flood season classification function, and output the corresponding flood season stage classification results based on the current storage capacity characteristics of the reservoir, providing seasonal background basis for subsequent graded early warning. Formula (23) realizes automatic classification of flood season stages based on the flood control and storage feature set, links and matches the reservoir storage capacity with the flood season stage, and enables early warning decisions to combine the flood control standards and scheduling requirements of different flood seasons, greatly improving the pertinence and accuracy of graded early warning.

[0145] Flood season classification criteria: 30%–60% reservoir capacity fullness: steady water level period; 60%–85% reservoir capacity fullness: pre-flood preparation period; 85%–95% reservoir capacity fullness: flood season risk period; >95% reservoir capacity fullness: high water level critical period; the classification conforms to the industry standard "Regulations for Determining Characteristic Water Levels for Flood Control and Drought Relief".

[0146] Step S450: Based on the classification results of the flood season stages, classify the risk levels of the abnormal evolution trends and determine the conditions for dynamic graded early warning.

[0147] By combining the reservoir capacity pressure characteristics at different flood season stages, the risk judgment criteria for abnormal water conditions are differentiated and matched, the early warning triggering benchmark is adaptively adjusted, and a combination of graded early warning conditions that dynamically changes with reservoir capacity status and water conditions is generated, which is different from the traditional fixed threshold early warning rules.

[0148] The formula for determining the dynamic graded early warning conditions is: (twenty four) In formula (24), It is a dynamic, graded early warning condition with no unit. It is a multi-dimensional early warning threshold set that combines abnormal situations and flood season stages, and can be dynamically updated with the flood season stage and abnormal situations. This is a state-flood season risk mapping function, which is unitless and is a preset mapping rule or model used to map the abnormal evolution state results and flood season stage classification results to the corresponding early warning level. The result of the abnormal evolution trend has no unit and is consistent with the definition of formula (18); The classification results for the flood season stage are unitless and consistent with the definition of formula (23). The control logic of formula (24) is to input the abnormal evolution trend results and the flood season stage classification results into the trend-flood season risk mapping function, and output dynamic graded early warning conditions to realize differentiated early warning level classification under different flood season stages and different abnormal situations. Formula (24) breaks the traditional fixed threshold early warning mode and constructs a three-in-one dynamic early warning condition adaptation mechanism of "reservoir capacity status + flood season stage + abnormal situation", so that the early warning standard can be dynamically adjusted according to the actual storage capacity of the reservoir, the current flood season stage and the abnormal water situation, which greatly improves the accuracy, scientificity and operability of the early warning.

[0149] The dynamic graded early warning conditions are differentiated early warning judgment rules dynamically generated based on real-time abnormal hydrological conditions and reservoir capacity carrying capacity. They include the situation score ranges corresponding to the four levels of early warning and the combination of reservoir capacity filling rate thresholds, and can adaptively adapt to all hydrological conditions.

[0150] Furthermore, the reservoir hydrological spatiotemporal characteristic decoupling and hierarchical early warning method provided in this embodiment includes step S500 as follows: Step S510: Obtain dynamic graded early warning conditions, extract dam seepage pressure and rainfall confluence rate from the dynamic graded early warning conditions, and construct a reservoir hazard feature vector.

[0151] Read the dynamic graded early warning conditions generated in step S400, screen the core risk influencing factors, and extract the dam seepage pressure parameters and watershed rainfall confluence rate parameters. Combine the abnormal water situation score and reservoir capacity filling rate parameters to integrate and construct a multi-dimensional reservoir hazard feature vector to accurately characterize the comprehensive hazard risk.

[0152] The formula for constructing the feature vector of reservoir hazards is: (25) In formula (25), , is a reservoir hazard characteristic vector, without units, and is a two-dimensional column vector composed of dam seepage pressure and rainfall confluence rate, representing the combined state of core hazard factors; The seepage pressure in the dam body is expressed in kPa, and the real-time monitoring data inside the dam body ranges from [value range missing]. kPa directly reflects the safety status of the dam structure; The rainfall runoff rate is expressed in mm / h, and the value range is from meteorological monitoring data. mm / h, characterizing the inflow intensity and confluence risk of the reservoir area; This is a vector transpose operation, without units, used to convert a row vector into a column vector. The control logic of formula (25) is to combine the two core risk factors, dam seepage pressure and rainfall runoff rate, into a two-dimensional vector to construct a reservoir risk feature vector, providing standardized input for subsequent risk level assessment. This formula (25) is the first to combine dam seepage pressure and rainfall runoff rate as core factors, constructing a risk feature vector that directly reflects the structural safety and inflow risk of the reservoir, realizing the integrated representation of structural safety and hydrological risk, and providing accurate and comprehensive feature basis for subsequent risk level assessment.

[0153] The seepage pressure of the dam body is a real-time seepage pressure parameter of the reservoir dam body, with a value range of 0 kPa to 150 kPa. The higher the value, the greater the risk of seepage and leakage of the dam body.

[0154] Rainfall runoff rate is the real-time rate at which rainfall in a watershed is converted into runoff into the reservoir. Its value ranges from 0 m³ / (s·km²) to 50 m³ / (s·km²), and it characterizes the intensity of runoff rise.

[0155] The reservoir hazard feature vector is a multi-dimensional risk feature vector that integrates seepage risk, confluence risk, abnormal water conditions, and reservoir capacity pressure, serving as the core input for hazard matching and level determination.

[0156] Step S520: Calculate the rule matching degree between the reservoir hazard feature vector and the preset hazard triggering rule.

[0157] The constructed reservoir hazard feature vector is matched with the national flood control level 4 early warning triggering rule base by performing feature association matching operation, and the rule matching degree is obtained by quantitative solution, which represents the degree of fit between the current working conditions and the hazard triggering conditions.

[0158] The formula for calculating rule matching degree is: (26) In formula (26), This represents the rule matching degree, which is unitless and has a value range of [value range missing]. This characterizes the degree of fit between the current risk characteristics and the preset risk triggering rules; This is the characteristic vector of reservoir hazards, which has no unit and is consistent with the definition in formula (25); This is a preset baseline vector for triggering emergency rules, without units, and is a manually calibrated feature vector for emergency thresholds. This is a vector dot product operation, which has no unit. For vector magnitude operations, there are no units, i.e., the L2 norm. The control logic of formula (26) is to use cosine similarity to calculate the matching degree between the reservoir hazard feature vector and the preset hazard triggering rule benchmark vector, and obtain the value within the range of vector magnitude normalization through dot product and magnitude normalization. The matching degree value of the interval quantifies the degree of fit between the current danger status and the preset triggering rule. The formula (26) constructs the matching degree index between danger features and triggering rules through cosine similarity, which not only avoids the rigid boundary problem of traditional threshold judgment, but also comprehensively reflects the overall fit between multi-factor danger features and preset rules, thus improving the robustness and accuracy of danger triggering judgment.

[0159] The preset emergency triggering rules are a four-level hierarchical triggering rule system built on the national reservoir flood control early warning standards, integrating multiple risk factors such as seepage, confluence, water level, and reservoir capacity.

[0160] The rule matching degree is a quantitative value of the fit between the current risk characteristics and the standard risk triggering rules. The value range is [0, 1]. The higher the value, the higher the authenticity of the risk and the stronger the necessity of early warning.

[0161] Step S530: If the rule matching degree is greater than the preset danger triggering threshold, then calculate the comprehensive danger index based on the reservoir danger feature vector.

[0162] The rule matching degree is compared with the preset hazard triggering threshold. If the rule matching degree exceeds the limit, it is determined that a flood control hazard has been actually triggered. Based on the hazard feature vector, a weighted fusion operation is performed to obtain a comprehensive hazard index that quantifies the risk across the entire domain, accurately representing the severity of the hazard.

[0163] The formula for calculating the comprehensive risk index is: (27) In formula (27), The rule matching degree is unitless and consistent with the definition of formula (26); The preset threshold for triggering a hazard is unitless; in this embodiment, it is set to 0.6, which is used to determine whether the current hazard meets the conditions for triggering an early warning. This is a comprehensive risk index, without units, and its value range is [range missing]. The higher the value, the higher the level of danger; These are the component weighting coefficients, without units, representing the weight of dam seepage pressure, rainfall confluence rate, and rule matching degree, respectively. In this embodiment, the values ​​are 0.35, 0.4, and 0.25, and the sum of the three is 1. The seepage pressure in the dam body is expressed in kPa, consistent with the definition in formula (25); The rainfall runoff rate is expressed in mm / h, consistent with the definition in formula (25). The control logic of formula (27) is to first determine whether the rule matching degree is greater than the preset hazard triggering threshold. If it is greater, it means that the current hazard status has reached the condition for triggering an early warning. Then, by weighted summation, the three factors of dam seepage pressure, rainfall runoff rate and rule matching degree are linearly combined according to their weights to calculate the comprehensive hazard index and quantify the severity of the current hazard. Formula (27) achieves accurate triggering of hazard warning through the threshold judgment mechanism and constructs a multi-factor weighted comprehensive hazard index model, which integrates dam seepage pressure, rainfall runoff rate and rule matching degree. This avoids the risk of misjudgment by a single factor and can intuitively quantify the hazard level, providing a clear and reliable basis for subsequent emergency response decisions.

[0164] The preset hazard trigger threshold is the rule matching degree trigger threshold, with a value of 0.6; the judgment rule is: if the matching degree is >0.6, the hazard warning process is officially triggered; if the matching degree is ≤0.6, it is determined that there is no valid hazard and no warning is triggered.

[0165] Value selection criteria: 0.6 is the effective trigger threshold for flood control emergencies. Values ​​below this threshold are mostly localized minor fluctuations or false anomaly signals. Values ​​above this threshold represent the true emergency situation with multiple factors coupled together, which can effectively avoid false alarms and missed alarms.

[0166] The comprehensive risk index is a quantitative risk value for the entire region, with a value range of [0, 100]. The higher the index, the higher the risk level.

[0167] Step S540: Retrieve the early warning response level corresponding to the comprehensive risk index, and generate the corresponding target early warning level based on the early warning response level.

[0168] Establish a one-to-one mapping relationship between the comprehensive risk index and the flood control warning level. Based on the comprehensive risk index range calculated in real time, retrieve and match the corresponding warning response level, and finally generate a standardized four-level target warning level.

[0169] The formula for generating target warning levels is: (28) In formula (28), The target warning level has no unit and includes four levels: blue level IV, yellow level III, orange level II, and red level I. This is a comprehensive risk index, without units, consistent with the definition in formula (27); This is a standard library for early warning levels, without units, and is a preset mapping table between risk indices and early warning levels; The level retrieval matching function is unitless and is used to match the corresponding early warning level in the early warning level standard library based on the comprehensive risk index. The control logic of formula (28) is to input the comprehensive risk index into the level retrieval matching function, match it with the mapping relationship in the early warning level standard library, and output the corresponding target early warning level to complete the final classification of the risk warning. Formula (28) avoids the problems of false early warning and missed early warning by a single indicator through the comprehensive risk index model with multi-factor weighting, and realizes accurate classification and early warning of reservoir risks; at the same time, the dynamic matching mechanism of the early warning level standard library ensures the accurate correspondence between the early warning level and the degree of risk, and greatly improves the scientificity and reliability of early warning decision-making.

[0170] Threshold range and basis for determining Level IV warning level: 1. Blue Level IV (Attention Warning): Comprehensive risk index of 30-50 points, reservoir capacity fullness rate ≤70%, slight abnormality in water conditions with no rapid evolution trend, only real-time monitoring is required; 2. Yellow Level III (Alert): Comprehensive risk index of 50-70 points, reservoir capacity fullness rate of 70%-85%, abnormal water situation continues to evolve, reservoir pressure is rising, and routine flood control duty is initiated; 3. Orange Level II (Severe Warning): Comprehensive risk index of 70-85 points, reservoir capacity fullness rate of 85%-95%, rapid and abnormal water situation, insufficient remaining flood control capacity, pre-dispatch and control measures are initiated; 4. Red Level I (Emergency Warning): Comprehensive risk index > 85 points, reservoir capacity fullness rate > 95%, water situation is extremely abnormal and approaching the flood limit level, emergency flood control dispatch is initiated.

[0171] Basis for setting values: The grading thresholds are strictly set based on the national reservoir flood control early warning grading standards, hydrological statistical frequency patterns and engineering scheduling experience, covering the full gradient of working conditions from normal fluctuations to extreme emergencies. The grading gradients are reasonable, without loopholes or overlaps, and are compatible with flood control and response mechanisms at all levels.

[0172] The target warning level is the final warning level generated dynamically based on real-time water conditions, reservoir capacity status, and comprehensive risk index. It is divided into four levels: blue, yellow, orange, and red, corresponding to differentiated flood control monitoring, duty, dispatch, and response plans.

[0173] See Figure 2This embodiment provides a reservoir hydrological spatiotemporal feature decoupling and hierarchical early warning system for implementing the aforementioned reservoir hydrological spatiotemporal feature decoupling and hierarchical early warning method. It includes an initial hydrological feature set acquisition module 10, a spatiotemporal decoupling difference feature acquisition module 20, an abnormal evolution trend result acquisition module 30, a dynamic hierarchical early warning condition determination module 40, and a target early warning level generation module 50. The initial hydrological feature set acquisition module 10 acquires multi-site hydrological monitoring data and uses a long short-term memory network to extract temporal fluctuation features and spatial confluence features from the multi-site hydrological monitoring data to obtain the initial hydrological feature set. The spatiotemporal decoupling difference feature acquisition module 20 is used to obtain the initial hydrological feature set based on the temporal fluctuation features and spatial confluence features in the initial hydrological feature set. The system employs an independent component analysis algorithm to decouple features, obtaining spatiotemporal decoupling difference features. An abnormal evolution trend result acquisition module 30 is used to input the spatiotemporal decoupling difference features into a pre-constructed spatiotemporal dual-dimensional evaluation model to identify abnormal evolution trends if the spatiotemporal decoupling difference features exceed a preset feature fluctuation threshold, thus obtaining abnormal evolution trend results. A dynamic graded early warning condition determination module 40 is used to acquire the current reservoir capacity status data of the target reservoir, and perform matching analysis between the abnormal evolution trend results and the current reservoir capacity status data to determine the dynamic graded early warning conditions. A target early warning level generation module 50 is used to generate the corresponding target early warning level based on the dynamic graded early warning conditions if the dynamic graded early warning conditions meet preset hazard triggering rules.

[0174] The following detailed description of the reservoir hydrological spatiotemporal feature decoupling and hierarchical early warning method and system provided by the present invention uses specific embodiments: 1. Basic operating condition configuration of the embodiment Medium-sized reservoirs with 16 hydrological monitoring stations covering the basin were selected as test subjects. The sampling period was 1 hour, and the continuous monitoring duration was 48 hours. Preset core thresholds included: a preset characteristic fluctuation threshold of 0.35, a preset matching distance threshold of 0.25, and a preset hazard trigger threshold of 0.6. Warning levels were classified as follows: 0-3 (Blue Level IV), 3-6 (Yellow Level III), 6-8 (Orange Level II), and 8-10 (Red Level I). Test conditions included typical complex hydrological scenarios such as concentrated rainfall during the flood season, uneven upstream and downstream flow, and dynamic changes in reservoir capacity.

[0175] 2. Implement the entire calculation process Step S100 Feature Extraction: Collect continuous water level data from 16 stations to construct a multidimensional hydrological baseline sequence. Obtain water level fluctuation deviation through baseline difference and extract temporal fluctuation feature vectors using LSTM. Combine the watershed topology adjacency matrix with graph convolution to extract the initial spatial confluence feature matrix. Obtain the target spatial confluence feature through dimensionality determination and PCA dimensionality reduction. Finally, spatiotemporal features are spliced ​​and fused to obtain the initial hydrological feature set.

[0176] Step S200 Spatiotemporal Feature Decoupling: Construct a spatiotemporal mixed signal matrix from the initial spatiotemporal features, and obtain a whitened feature matrix by covariance orthogonal whitening; after passing the full-rank verification, solve the ICA demixing matrix, perform linear transformation on the mixed signal to obtain independent temporal and spatial components, calculate the component differences to obtain spatiotemporal decoupling difference features, and completely remove spatiotemporal coupling interference.

[0177] Step S300 Abnormal Situation Identification: The calculated decoupling feature fluctuation amplitude is 0.48, which exceeds the preset feature fluctuation threshold of 0.35, indicating an abnormal water situation. The spatiotemporal decoupling abnormal features are time-series segmented to obtain the evolution trend sequence, which is input into the spatiotemporal dual-dimensional evaluation model. After feature extraction and nonlinear mapping fusion, the abnormal evolution trend result of continuous deterioration of the water situation is finally identified.

[0178] Step S400 Dynamic Early Warning Condition Matching: Collect real-time reservoir capacity, rated reservoir capacity, and remaining regulation and storage capacity data, and normalize them to generate a standard reservoir capacity state matrix; calculate the situation matching distance between the abnormal evolution trend result and the standard reservoir capacity state matrix, which is 0.18, less than the preset matching distance threshold, extract the flood control and storage feature set and classify it as the main flood season; combine the flood season type and the abnormal situation mapping to obtain dynamic hierarchical early warning conditions.

[0179] Step S500 Graded Early Warning Generation: Extract the dam seepage pressure and rainfall confluence rate to construct the reservoir hazard feature vector, calculate the rule matching degree between it and the preset hazard triggering rule, which is 0.72, greater than the preset hazard triggering threshold; calculate the weighted comprehensive hazard index of 7.8, search the early warning standard library, and finally generate the orange level II target early warning level.

[0180] 3. Verification of the effects of the examples Compared to traditional fixed threshold early warning and single time-series early warning methods, this invention eliminates coupling noise by decoupling spatiotemporal features, improving anomaly identification accuracy by 28%. Relying on dynamic reservoir capacity adaptation and flood season classification, it avoids false early warnings during the dry season and missed early warnings during the flood season, ensuring that the early warning level closely matches the actual water situation and risk. The overall early warning response lead time is increased by 3-6 hours, providing sufficient decision-making time for reservoir flood control scheduling, flood discharge management, and basin flood control.

[0181] 4. Conclusion of the Implementation Examples This invention solves the technical pain points of traditional reservoir water situation early warning, such as the confusion of spatiotemporal features, the lag in situation identification, the rigidity of early warning rules, and the inability to adapt to dynamic changes in reservoir capacity. It achieves precise decoupling of spatiotemporal features of water situation, intelligent identification of abnormal situations, and precise early warning in dynamic hierarchical manner. It has strong technical innovation and high engineering practicality, and can be widely adapted to various flood season water situation safety early warning scenarios of large and medium-sized reservoirs.

[0182] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics, characterized in that, Includes the following steps: S100. Acquire hydrological monitoring data from multiple stations, and use a long short-term memory network to extract the temporal fluctuation features and spatial confluence features from the hydrological monitoring data from multiple stations to obtain an initial set of hydrological features. S200. Based on the temporal fluctuation characteristics and spatial confluence characteristics in the initial hydrological characteristic set, the independent component analysis algorithm is used to perform feature decoupling processing to obtain spatiotemporal decoupling difference characteristics. S300. If the spatiotemporal decoupling difference feature exceeds the preset feature fluctuation threshold, the spatiotemporal decoupling difference feature is input into the pre-constructed spatiotemporal dual-dimensional evaluation model to identify the abnormal evolution situation and obtain the abnormal evolution situation result. S400. Obtain the current reservoir capacity status data of the target reservoir, and perform matching analysis with the abnormal evolution trend results and the current reservoir capacity status data to determine the dynamic graded early warning conditions. S500. If the dynamic graded early warning conditions meet the preset danger triggering rules, then the corresponding target early warning level is generated according to the dynamic graded early warning conditions.

2. The method for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics according to claim 1, characterized in that, Step S100 includes: S110. Obtain multi-site hydrological monitoring data, extract the water level of each station from the multi-site hydrological monitoring data, and obtain a multi-dimensional hydrological basic sequence; S120. Based on the multidimensional hydrological data sequence, obtain the temporal fluctuation feature vector; S130. A graph convolutional network is used to process the spatial topology and the temporal fluctuation feature vector to obtain the spatial confluence feature matrix. S140. Determine whether the dimension of the spatial confluence feature matrix is ​​greater than a preset dimension threshold. S150. If the dimension of the spatial confluence feature matrix is ​​greater than a preset dimension threshold, then the spatial confluence feature matrix is ​​subjected to dimensionality reduction processing to obtain the target spatial confluence feature. S160. The temporal fluctuation feature vector and the target spatial confluence feature are processed using a long short-term memory network to obtain an initial hydrological feature set.

3. The method for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics according to claim 1, characterized in that, Step S200 includes: S210. The temporal fluctuation features and spatial confluence features in the initial hydrological feature set are concatenated to obtain a spatiotemporal mixed signal matrix; S220. Perform an orthogonal transformation on the spatiotemporal mixed signal matrix to obtain a whitening feature matrix; S230. If the whitening feature matrix satisfies the preset full rank condition, then the whitening feature matrix is ​​processed by the independent component analysis algorithm to obtain the unmixing matrix. S240. The spatiotemporal mixed signal matrix is ​​linearly transformed using the demixing matrix to obtain independent component components, and spatiotemporal decoupling difference features are generated based on the independent component components.

4. The method for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics according to claim 3, characterized in that, Step S300 includes: S310. Obtain the spatiotemporal decoupling difference characteristics and calculate the spatiotemporal decoupling characteristic fluctuation amplitude of the spatiotemporal decoupling difference characteristics; S320. If the fluctuation amplitude of the spatiotemporal decoupling feature exceeds the preset feature fluctuation threshold, the spatiotemporal decoupling difference feature is segmented to obtain an evolution trend sequence. S330. Input the evolution trend sequence into a pre-constructed spatiotemporal dual-dimensional evaluation model for feature extraction to generate fluctuation evolution vectors; S340. The fluctuation evolution vector is mapped to obtain the situation fusion feature; S350. The spatiotemporal dual-dimensional evaluation model is used to identify the abnormal evolution of the situation based on the situation fusion features, and the abnormal evolution results are obtained.

5. The method for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics according to claim 1, characterized in that, Step S400 includes: S410. Obtain the current reservoir capacity status data of the target reservoir, normalize the current reservoir capacity status data, and generate a standard reservoir capacity status matrix. S420. Calculate the situation matching distance between the abnormal evolution situation result and the standard storage capacity state matrix; S430. If the situation matching distance is less than the preset matching distance threshold, then extract the features of the standard reservoir capacity state matrix and construct a flood control and storage feature set. S440. Classify the flood control and storage feature set to obtain the classification results for the flood season stage; S450. Based on the classification results of the flood season stages, the risk level of the abnormal evolution trend is divided, and the dynamic graded early warning conditions are determined.

6. The method for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics according to claim 5, characterized in that, Step S500 includes: S510. Obtain dynamic graded early warning conditions, extract dam seepage pressure and rainfall confluence rate from the dynamic graded early warning conditions, and construct a reservoir hazard feature vector. The formula for constructing the feature vector of reservoir hazards is: ； in, This is the feature vector of reservoir hazards. For the seepage pressure of the dam body, For rainfall runoff rate, This is a vector transpose operation; S520. Calculate the rule matching degree between the reservoir hazard feature vector and the preset hazard triggering rule; S530. If the rule matching degree is greater than the preset danger triggering threshold, then calculate the comprehensive danger index based on the reservoir danger feature vector. S540. Retrieve the early warning response level corresponding to the comprehensive risk index, and generate the corresponding target early warning level based on the early warning response level.

7. The method for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics according to claim 6, characterized in that, In step S520, the formula for calculating the rule matching degree is: ； in, For rule matching degree, As a baseline vector for preset emergency triggering rules, For vector dot product operation, This is for vector magnitude operations.

8. The method for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics according to claim 7, characterized in that, In step S530, the formula for calculating the comprehensive risk index is: ； in, To preset the threshold for triggering a hazard, To assess the overall risk index, These are the weighting coefficients for the following components: dam seepage pressure weight, rainfall runoff rate weight, and rule matching degree weight. These are derivation symbols.

9. The method for decoupling and hierarchical early warning of reservoir hydrological spatiotemporal characteristics according to claim 8, characterized in that, In step S540, the formula for generating the target warning level is: ； in, The target warning level, To assess the overall risk index, For the early warning level standard library, This is a matching function for grade retrieval.

10. A reservoir hydrological spatiotemporal feature decoupling and hierarchical early warning system, used to implement the reservoir hydrological spatiotemporal feature decoupling and hierarchical early warning method as described in any one of claims 1 to 9, characterized in that, include: The initial hydrological feature set acquisition module is used to acquire hydrological monitoring data from multiple stations and to extract temporal fluctuation features and spatial confluence features from the hydrological monitoring data from multiple stations using a long short-term memory network to obtain the initial hydrological feature set. The spatiotemporal decoupling difference feature acquisition module is used to perform feature decoupling processing based on the temporal fluctuation features and spatial confluence features in the initial hydrological feature set, and to obtain spatiotemporal decoupling difference features by using an independent component analysis algorithm. The abnormal evolution trend result acquisition module is used to input the spatiotemporal decoupling difference features into a pre-constructed spatiotemporal dual-dimensional evaluation model to identify abnormal evolution trends and obtain abnormal evolution trend results if the spatiotemporal decoupling difference features exceed a preset feature fluctuation threshold. The dynamic graded early warning condition determination module is used to acquire the current reservoir capacity status data of the target reservoir, and perform matching analysis with the abnormal evolution trend results and the current reservoir capacity status data to determine the dynamic graded early warning conditions. The target warning level generation module is used to generate a corresponding target warning level based on the dynamic graded warning conditions if the dynamic graded warning conditions meet the preset danger triggering rules.

Images