Hydrological forecasting method and system based on adaptive correction chain

Abstract

The application provides a hydrological prediction method and system based on an adaptive correction chain, relates to the field of data processing, and comprises the following steps: generating a first reservoir inflow prediction result according to rainfall deviation traceability-dynamic correction chain and rainfall observation data; determining the multi-dimensional characteristics of a current scene, generating a second reservoir inflow prediction result according to the multi-dimensional characteristics of the current scene and historical similar sample matching-law application chain; generating a third reservoir inflow prediction result according to runoff generation and concentration knowledge-threshold constraint chain and underlying surface conditions, rainfall spatial distribution and basin state; and determining the dynamic weights of the first reservoir inflow prediction result, the second reservoir inflow prediction result and the third reservoir inflow prediction result to generate a final reservoir inflow prediction result, thereby improving the accuracy of the reservoir inflow prediction result.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to a hydrological forecasting method and system based on an adaptive correction chain. Background Technology

[0002] Hydrological forecasting is a quantitative or qualitative prediction of hydrological conditions (such as water level, flow, and water quality) for a certain period in the future, based on previous or current hydrological and meteorological information. It is a key basis for flood control and disaster reduction, rational utilization and management of water resources, scheduling of water conservancy projects, and ecological protection.

[0003] Traditional methods, such as time series analysis, regression analysis, and empirical models, are largely based on the assumption that "history repeats itself." They construct models for forecasting by exploring the statistical relationships between historical meteorological data (such as rainfall and temperature) and historical runoff data. However, their limitations are extremely prominent. On the one hand, when faced with "black swan" events such as torrential rains or extreme droughts that exceed historical records, the forecasts often deviate significantly from reality because the models have not learned such patterns, resulting in a substantial decrease in reliability. On the other hand, these models are essentially mathematical "black boxes" or "grey boxes," focusing only on the correlation between inputs and outputs while ignoring the intrinsic physical processes of hydrology (such as infiltration, evapotranspiration, and runoff timing), making it difficult to explain the forecast results and lacking mechanistic persuasiveness. Furthermore, when the underlying surface conditions of a watershed change due to human activities (such as land use change and water conservancy projects) or natural changes (such as landslides and debris flows), the original statistical relationships no longer apply. The models need to be recalibrated or even rebuilt, a cumbersome process with poor adaptability.

[0004] While conceptual hydrological models attempt to approximate runoff generation and confluence processes using simplified conceptual tanks or functions, representing an improvement over purely statistical models, they still have limitations. These models contain numerous parameters (such as evaporation coefficient, infiltration rate, and storage capacity) that need to be inferred from historical data. In watersheds with complex topography and strong spatial heterogeneity, a single set of parameters is insufficient to represent the entire watershed, and unit-by-unit calculations require substantial data support, making parameter calibration difficult. Furthermore, multiple combinations of different parameters may yield similar simulation results, i.e., parameters are "different but equally effective," making it difficult to determine a "unique optimal solution," thus reducing the model's physical consistency and the reliability of extrapolation forecasts. Simultaneously, model accuracy heavily depends on the accuracy and spatial distribution of input data (especially rainfall). For areas with sparse meteorological stations, such as mountainous areas and hydropower stations, station data may not represent the true spatial distribution of rainfall, leading to significant uncertainty in the model input.

[0005] Therefore, there is a need to provide hydrological forecasting methods and systems based on adaptive correction chains to improve the accuracy of reservoir inflow forecasts. Summary of the Invention

[0006] This invention provides a hydrological forecasting method based on an adaptive correction chain, comprising: acquiring rainfall observation data; generating a first reservoir inflow forecast result based on the rainfall deviation source tracing-dynamic correction chain and the rainfall observation data; determining the multi-dimensional features of the current scenario; generating a second reservoir inflow forecast result based on the historical similar sample matching-regular application chain and the multi-dimensional features of the current scenario; acquiring underlying surface conditions, rainfall spatial distribution, and watershed status; generating a third reservoir inflow forecast result based on the runoff generation and confluence knowledge-threshold constraint chain and the underlying surface conditions, rainfall spatial distribution, and watershed status; determining the dynamic weights of the first, second, and third reservoir inflow forecast results; and generating the final reservoir inflow forecast result based on the dynamic weights and strategy selection tree of the first, second, and third reservoir inflow forecast results.

[0007] Furthermore, based on the rainfall deviation source-dynamic correction chain and rainfall observation data, the first reservoir inflow forecast result is generated, including: aligning the observation data with the forecast grid using a spatiotemporal interpolation algorithm, calculating the areal rainfall error sequence, and determining the error statistical characteristics; determining the correction parameters based on the error statistical characteristics; generating the initial reservoir inflow forecast result using the first prediction model; and correcting the initial reservoir inflow forecast result based on the correction parameters to generate the first reservoir inflow forecast result.

[0008] Furthermore, based on error statistical characteristics, correction parameters are determined, including: determining error patterns based on real-time weather data and a weather-error pattern rule base; obtaining dynamic-thermal parameters of the current scenario; determining similar scenarios based on error statistical characteristics, error patterns, and dynamic-thermal parameters of the current scenario; and determining correction parameters based on similar scenarios.

[0009] Furthermore, the multi-dimensional features of the current scenario include rainfall characteristics, watershed status, and weather conditions; based on the historical similar sample matching-pattern application chain and the multi-dimensional features of the current scenario, a second reservoir inflow forecast result is generated, including: determining similar samples based on the multi-dimensional features of the current scenario, obtaining key hydrological feature values ​​of the similar samples; and generating the second reservoir inflow forecast result based on the key hydrological feature values ​​of the similar samples using a second prediction model.

[0010] Furthermore, based on the knowledge-threshold constraint chain of runoff generation and confluence, underlying surface conditions, spatial distribution of rainfall, and watershed status, the inflow forecast result for the third reservoir is generated, including: determining the optimal runoff generation parameters through real-time weather data and underlying surface conditions; generating the unit flow line based on the spatial distribution of rainfall and watershed status; and generating the inflow forecast result for the third reservoir based on the optimal runoff generation parameters and the unit flow line through the third forecast model, wherein the third forecast model embeds a physical constraint loss function.

[0011] Furthermore, the dynamic weights of the inflow forecasts for the first, second, and third reservoirs are determined, including: determining the basic weights of the inflow forecasts for the first, second, and third reservoirs based on historical forecast errors from the rainfall deviation source tracing-dynamic correction chain, historical similar sample matching-pattern application chain, and runoff generation and confluence knowledge-threshold constraint chain; and adjusting the basic weights of the inflow forecasts for the first, second, and third reservoirs based on the current physical mechanism to determine their dynamic weights.

[0012] Furthermore, based on the current physical mechanism, the basic weights of the inflow forecasts for the first, second, and third reservoirs are adjusted to determine their dynamic weights. This includes: generating a similarity enhancement factor for the first reservoir inflow forecast based on the similarity of similar scenarios; generating a similarity enhancement factor for the second reservoir inflow forecast based on the similarity of similar samples; determining a time decay factor for generating the second reservoir inflow forecast based on the time interval of similar samples; and determining a dynamic weight based on the current physical mechanism. The fit of the inflow forecasts for the first, second, and third reservoirs is determined. Based on the similarity enhancement factor, time decay factor, and fit of the inflow forecasts for the first, second, and third reservoirs, the basic weights of the inflow forecasts for the first, second, and third reservoirs are adjusted to determine their dynamic weights.

[0013] Furthermore, based on the dynamic weights and strategy selection tree of the inflow forecasts for the first, second, and third reservoirs, the final reservoir inflow forecast result is generated. This includes: determining a fusion mode based on the dynamic weights of the inflow forecasts for the first, second, and third reservoirs, wherein the fusion mode can be an optimal mode, a segmented mode, or a weighted average mode; and fusing the inflow forecasts for the first, second, and third reservoirs based on the fusion mode to generate the final reservoir inflow forecast result.

[0014] Furthermore, based on the dynamic weights of the inflow forecasts for the first, second, and third reservoirs, a fusion mode is determined, including: if the dynamic weight of any of the three reservoir inflow forecasts is greater than a preset weight threshold, then the fusion mode is the optimal mode; if the dynamic weight of any of the three reservoir inflow forecasts is less than or equal to a preset weight threshold, then based on the river network humidity index and the spatial heterogeneity of rainfall, it is determined whether the fusion mode is a segmented mode.

[0015] This invention provides a hydrological forecasting system based on an adaptive correction chain. Applying the aforementioned hydrological forecasting method based on an adaptive correction chain, it includes: a first forecasting module for acquiring rainfall observation data and generating a first reservoir inflow forecast result based on the rainfall deviation source tracing-dynamic correction chain and the rainfall observation data; a second forecasting module for determining the multi-dimensional features of the current scenario and generating a second reservoir inflow forecast result based on the historical similar sample matching-regular application chain and the multi-dimensional features of the current scenario; and a third forecasting module for acquiring underlying surface conditions, Based on the spatial distribution of rainfall and the watershed status, and according to the knowledge chain of runoff generation and confluence, threshold constraint chain, underlying surface conditions, and the spatial distribution of rainfall and the watershed status, the inflow forecast result of the third reservoir is generated. The forecast fusion module is used to determine the dynamic weights of the inflow forecast results of the first, second, and third reservoirs, and based on the dynamic weights of the inflow forecast results of the first, second, and third reservoirs and the strategy selection tree, the final reservoir inflow forecast result is generated.

[0016] Compared with existing technologies, the hydrological forecasting method and system based on adaptive correction chains provided by this invention have at least the following beneficial effects:

[0017] 1. Utilizing the rainfall deviation source tracing-dynamic correction chain, a first forecast result is generated by combining rainfall observation data, ensuring the rationality of the basic forecast from the perspective of rainfall deviation; through the historical similar sample matching-pattern application chain, a second forecast result is generated based on the multi-dimensional characteristics of the current scenario, improving forecast reliability by leveraging historical patterns; using the runoff generation and confluence knowledge-threshold constraint chain, a third forecast result is generated based on multiple factors such as the underlying surface, integrating professional knowledge to ensure the scientific nature of the forecast, and comprehensively improving forecast accuracy by integrating multi-source information.

[0018] 2. By determining the dynamic weights of the three forecast results, the proportion of each result in the final forecast can be flexibly adjusted based on real-time changes in data and scenarios. This dynamic allocation mechanism enables the forecasting method to adapt to different hydrological scenarios, responding promptly to changes in rainfall characteristics or underlying surface conditions, ensuring that the forecast results always closely match the actual situation.

[0019] 3. The final forecast results are generated based on dynamic weights and a strategy selection tree. The strategy selection tree comprehensively considers multiple factors and logical relationships to optimize the forecast results under different weight combinations. This further improves the accuracy and reliability of the forecast, providing a more scientific and accurate basis for decisions such as reservoir scheduling. Attached Figure Description

[0020] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0021] Figure 1 This is a flowchart illustrating a hydrological forecasting method based on an adaptive correction chain, according to some embodiments of this specification.

[0022] Figure 2 This is a schematic diagram of a hydrological forecasting system based on an adaptive correction chain, as shown in some embodiments of this specification. Detailed Implementation

[0023] To more clearly illustrate the technical solutions of the embodiments in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of this specification. For those skilled in the art, these drawings can be applied to other similar scenarios without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0024] Figure 1 This is a flowchart illustrating a hydrological forecasting method based on an adaptive correction chain, as shown in some embodiments of this specification. Figure 1As shown, the hydrological forecasting method based on adaptive correction chains may include the following steps.

[0025] Step 110: Obtain rainfall observation data, and generate the inflow forecast result of the first reservoir based on the rainfall deviation source-dynamic correction chain and rainfall observation data.

[0026] Among them, rainfall observation data can be obtained from real-time ground rain gauges, radar quantitative precipitation estimates, etc.

[0027] Specifically, it includes:

[0028] The observation data is aligned with the forecast grid using a spatiotemporal interpolation algorithm. The areal rainfall error sequence is calculated to determine the statistical characteristics of the error. The areal rainfall error sequence refers to the sequence of precipitation error over time at each grid point or sub-region within a certain region or the entire forecast region, calculated by comparing the observation data with the numerical weather prediction grid data and aligning them through spatiotemporal interpolation. The statistical characteristics of the error may include mean, variance, standard deviation, systematic overestimation or underestimation trend, random fluctuation, etc.

[0029] Based on the statistical characteristics of the error, the correction parameters are determined;

[0030] The first prediction model is used to generate the initial reservoir inflow forecast results;

[0031] The initial reservoir inflow forecast is corrected based on the correction parameters to generate the first reservoir inflow forecast.

[0032] In some embodiments, determining correction parameters based on error statistical characteristics includes:

[0033] Based on real-time weather data and a weather-error pattern rule base, the error pattern is determined.

[0034] Obtain the dynamic-thermal parameters of the current scene (e.g., vertical velocity, water vapor flux divergence, convective available potential energy, etc.);

[0035] Based on error statistics, error patterns, and dynamic-thermal parameters of the current scenario, similar scenarios are identified.

[0036] Based on similar scenarios, determine the correction parameters.

[0037] Specifically, a "weather type-error pattern" rule base is pre-constructed based on historical data and meteorological knowledge. This rule base contains the relationships between different weather types and their corresponding error patterns. Meteorological elements and characteristics extracted from real-time weather data are matched against the weather types in the rule base. For example, if real-time data shows characteristics of warm-sector heavy rainfall, it is matched against the "warm-sector heavy rainfall" weather type in the rule base. Once a matching weather type is found, the system can determine the corresponding error pattern based on the pre-defined relationships in the rule base. For instance, for the "warm-sector heavy rainfall" weather type, the rule base might indicate an error pattern that tends to underestimate precipitation.

[0038] Correction parameters (e.g., scaling factor, spatial adjustment vector, time adjustment factor, etc.) for different sample scenarios are pre-constructed. For each sample scenario, the similarity between the sample scenario and the current scenario is calculated, including the error statistics, error pattern, and dynamic-thermal parameters of the current scenario. The sample scenario with the highest similarity is taken as the similar scenario, and the correction parameters of the similar scenario are retrieved as the correction parameters of the current scenario.

[0039] The first prediction model is a mathematical model built based on historical data, weather forecasts, and watershed characteristics, aiming to predict reservoir inflow over a future period. The model's construction process includes data collection, feature extraction, model training, and validation. By collecting historical inflow data, weather forecast data (such as rainfall and evaporation), and watershed characteristic data (such as topography and soil type), key features are extracted, and the first prediction model is trained using machine learning or statistical methods. Finally, the model's predictive performance is evaluated using a validation set. Current weather forecast data and watershed status data are input into the first prediction model. Based on the input data, the model uses its internal algorithms and parameters to calculate the reservoir inflow forecast for a future period. This result is a preliminary, uncorrected prediction. A scaling factor can be applied to correct for magnitude biases in the initial forecast. For example, if initial forecasts are generally low, a scaling factor greater than 1 can improve the forecast value. A spatial adjustment vector can be applied to correct for spatial distribution biases in the initial forecast. For example, if forecast values ​​are too high for some areas and too low for others, the forecast values ​​for each area can be adjusted to better reflect the actual distribution. Time adjustment factors can be applied to correct for deviations in the temporal variation of the initial forecast results. For instance, if the predicted rainfall event is earlier or later than expected, the time adjustment factor can be adjusted to make the forecast results more consistent with actual time variations. By comprehensively applying correction parameters such as the proportional coefficient, spatial adjustment vector, and time adjustment factor to the initial forecast results, a comprehensive correction is performed, resulting in the predicted inflow to the first reservoir. This result is more accurate and reliable than the initial forecast, and better reflects the inflow situation to the reservoir over a future period.

[0040] Step 120: Determine the multi-dimensional features of the current scene, and generate the inflow forecast result of the second reservoir based on the historical similar sample matching-pattern application chain and the multi-dimensional features of the current scene.

[0041] The current scenario features multiple dimensions, including rainfall characteristics, watershed status, and weather conditions. For example, rainfall characteristics include previous cumulative rainfall, the spatiotemporal distribution of forecast rainfall, and the location of the rainfall center; watershed status includes rising flow, soil saturation, and baseflow level; and weather conditions include the type of influencing system and water vapor transport conditions.

[0042] In some embodiments, based on historical similar sample matching-pattern application chains and multi-dimensional features of the current scenario, a second reservoir inflow forecast result is generated, including:

[0043] Based on the multi-dimensional features of the current scenario, similar samples are identified, and key hydrological feature values ​​of similar samples are obtained. Specifically, by using a hybrid similarity measurement algorithm that integrates Euclidean distance, dynamic time warping, and expert weight constraints, the comprehensive similarity between the current forecast scenario and sample cases in terms of rainfall process, flow process, and weather background is quantified. Based on the comprehensive similarity, Top-K similar samples are intelligently retrieved from the sample case library, such as the top 3, top 5, etc., and the key hydrological feature values ​​of similar samples, such as the flow rate increase, peak time, and flood volume at that time, are output as correction references.

[0044] The second prediction model generates the inflow forecast results for the second reservoir based on the key hydrological feature values ​​of similar samples.

[0045] Specifically, the second prediction model can employ machine learning models (such as Long Short-Term Memory Networks, Transformers, etc.) or statistical models (such as time series analysis models, etc.). Based on key hydrological features of similar samples and current inflow rates at multiple time points, the second prediction model generates a forecast of the inflow rate to the second reservoir.

[0046] Step 130: Obtain the underlying surface conditions, spatial distribution of rainfall, and watershed status. Based on the knowledge of runoff generation and confluence - threshold constraint chain and the underlying surface conditions, spatial distribution of rainfall, and watershed status, generate the inflow forecast results for the third reservoir.

[0047] Specifically, it includes:

[0048] Optimal runoff generation parameters are determined by using real-time weather data (e.g., rainfall intensity, rainfall duration) and underlying surface conditions (e.g., soil moisture, vegetation interception, land use).

[0049] Based on the spatial distribution of rainfall and the state of the watershed, a unit line of convergence is generated;

[0050] The third forecast model generates the inflow forecast of the third reservoir based on the optimal runoff parameters and the unit flow line. The third forecast model embeds a physical constraint loss function.

[0051] Specifically, a runoff generation parameter database can be constructed, which can store optimal runoff generation parameters (e.g., water storage capacity, infiltration rate, etc.) for multiple cases. Different cases correspond to different rainfall types (e.g., typhoon rainstorms, frontal rainstorms, etc.) and underlying surface conditions. By using real-time weather data, the current rainfall type is determined. Multiple cases corresponding to the current rainfall type are searched from the runoff generation parameter database. The similarity between the underlying surface conditions of the current scenario and the underlying surface conditions of the cases corresponding to the current rainfall type is calculated. The optimal runoff generation parameter corresponding to the case with the highest similarity is taken as the optimal runoff generation parameter for the current scenario.

[0052] Spatial distribution of rainfall can include at least the following data:

[0053] The spatial coordinates of the maximum rainfall intensity are extracted from the center of the rainfall using a two-dimensional Gaussian fitting or centroid algorithm.

[0054] Rainfall intensity gradient: Calculate the distance attenuation coefficient between each point in the rainfall field and the rainfall center to reflect spatial heterogeneity.

[0055] The direction and speed of rainfall movement are tracked by continuous radar data to determine the displacement of the rainfall center and quantify its dynamic impact on the confluence path.

[0056] The watershed status can include river network humidity index, soil moisture, topographic slope, land use type, etc.

[0057] The third forecast model employs a CNN-LSTM architecture. The radar-retrieved rainfall field (a two-dimensional matrix) is input into the CNN, and spatial features (such as rainfall center location and intensity gradient) are extracted through convolutional and pooling layers. The output feature map encodes key patterns in the spatial distribution of rainfall. The spatial features output by the CNN, along with watershed state parameters, are input into the LSTM in a time series manner to capture the impact of dynamic processes such as rainfall movement and changes in river network humidity index on confluence. LSTM units retain long-term dependencies (such as the cumulative effect of continuous rainfall center movement on the confluence path) through a memory gating mechanism. The output is either time-varying unitline shape parameters (such as peak time, peak flow, and time history) or a directly generated unitline function. For example, when rainfall is concentrated in the near-channel area (CNN identifies rainfall centers close to the river network) and the river network humidity index is high (LSTM captures the storage effect), the third forecast model automatically shortens the peak time and reduces the peak flow, reflecting the physical process of accelerated confluence but peak reduction through storage.

[0058] Traditional unit hydrographs (UH) rely on fixed assumptions (such as uniform rainfall distribution and constant watershed conditions), making it difficult to reflect the impact of spatial heterogeneity of rainfall and dynamic watershed responses (such as river regulation and soil moisture changes) on the runoff process. By integrating real-time rainfall spatial distribution and watershed conditions, this paper abandons the linear time-invariant characteristics of traditional unit hydrographs, generates time-varying unit hydrographs that match the real-time scenario, and improves the forecast accuracy of reservoir inflow.

[0059] Step 140: Determine the dynamic weights of the inflow forecast results for the first reservoir, the second reservoir, and the third reservoir.

[0060] Specifically, it includes:

[0061] Based on the historical forecast errors of the rainfall deviation source tracing-dynamic correction chain, historical similar sample matching-regulation application chain, and runoff generation and confluence knowledge-threshold constraint chain, the basic weights of the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir are determined.

[0062] Based on the current physical mechanism, the basic weights of the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir are adjusted to determine the dynamic weights of the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir.

[0063] Specifically, the basic weights of the inflow forecasts for the first, second, and third reservoirs can be calculated using the following formula:

[0064]

[0065] in, The base weights for the results corresponding to the m-th reservoir inflow forecasting method are: Let m be the historical average relative error corresponding to the m-th method for predicting reservoir inflow. Let N be the historical average relative error corresponding to the nth method for forecasting reservoir inflow, and N be the total number of methods for forecasting reservoir inflow, i.e., 3.

[0066] In some embodiments, based on the physical mechanism of the current state, the basic weights of the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir are adjusted to determine the dynamic weights of the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir, including:

[0067] Based on the similarity of similar scenarios, a similarity enhancement factor is generated for the forecast results of the inflow of the first reservoir. The greater the similarity of the similar scenarios, the greater the similarity enhancement factor for the forecast results of the inflow of the first reservoir.

[0068] Based on the similarity of similar samples, a similarity enhancement factor is generated for the forecast results of the inflow into the second reservoir. Based on the time interval of similar samples, a time decay factor is determined for the forecast results of the inflow into the second reservoir. The shorter the time interval between the similar sample and the current time, the larger the time decay factor of the forecast results of the inflow into the second reservoir.

[0069] Based on the physical mechanism of the current state, determine the fit of the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir.

[0070] Based on the similarity enhancement factor of the inflow forecast results of the first reservoir, the similarity enhancement factor and time decay factor of the inflow forecast results of the second reservoir, and the fit of the inflow forecast results of the first, second and third reservoirs, the basic weights of the inflow forecast results of the first, second and third reservoirs are adjusted to determine the dynamic weights of the inflow forecast results of the first, second and third reservoirs.

[0071] Specifically, the weights of the inflow forecast results for the second reservoir can be adjusted based on the time decay factor of the inflow forecast results, according to the following formula:

[0072]

[0073] in, λ represents the weight of the second reservoir inflow forecast results after adjustment by the time decay factor based on the second reservoir inflow forecast results, where Δt is the time interval between similar samples and λ is the decay factor.

[0074] The weights of the inflow forecast results for the second reservoir can be adjusted based on the similarity enhancement factor of the inflow forecast results, according to the following formula:

[0075]

[0076] in, The dynamic weights of the inflow forecast results for the second reservoir are used. This is a similarity enhancement factor for the forecast results of the inflow into the second reservoir.

[0077] Step 150: Based on the dynamic weights and strategy selection tree of the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir, generate the final reservoir inflow forecast result.

[0078] Specifically, it includes:

[0079] Based on the dynamic weights of the inflow forecasts for the first, second, and third reservoirs, a fusion mode is determined, which may be an optimal mode, a segmented mode, or a weighted average mode.

[0080] Based on the fusion model, the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir are merged to generate the final reservoir inflow forecast result.

[0081] In some embodiments, a fusion mode is determined based on the dynamic weights of the inflow forecasts for the first, second, and third reservoirs, including:

[0082] If the dynamic weight of the inflow forecast results of the first reservoir, the second reservoir, or the third reservoir is greater than the preset weight threshold (e.g., 0.7), the fusion mode is the optimal mode. For example, if the dynamic weight of the inflow forecast result of the first reservoir is significantly higher than the other two results, then the inflow forecast result of the first reservoir will be directly used as the final inflow forecast result of the reservoir.

[0083] If the dynamic weights of the predicted inflows to the first, second, and third reservoirs are less than or equal to a preset weight threshold, then based on the river network humidity index and the spatial heterogeneity of rainfall, it is determined whether the fusion model is a segmented model. Specifically, when the river network humidity index is high, the propagation and evolution of floods may become more complex due to the influence of river channel regulation. In this case, it is necessary to consider the flood characteristics at different stages to select an appropriate forecast fusion scheme. When rainfall is concentrated in certain areas, these areas may generate surface runoff more quickly, thus affecting the flood formation and evolution process of the entire basin. Therefore, when there is significant spatial heterogeneity in rainfall, different dominant schemes need to be adopted according to different forecast stages. As an example, if the river network humidity index is greater than 2 and the spatial heterogeneity of rainfall is greater than 1.5, then the fusion model is a segmented model.

[0084] Specifically, under the scenario segmentation and fusion strategy, the forecast period is divided, and different dominant schemes are adopted in different stages.

[0085] In the lead-up to forecasts of concentrated rainfall and dramatic rises in water levels, the forecast of the inflow into the first reservoir, which is highly sensitive to input changes, is used because rainfall at this stage has a significant impact on flow rate changes. This is because the initial rainfall directly determines the initial increase in water volume, and the forecast of the inflow into the first reservoir better reflects these earlier flow rate variations.

[0086] During the peak flood season, the inflow forecast from the third reservoir, which is more physically based, will be adopted. The third reservoir's inflow forecast is based on physical principles and can more accurately describe the formation and propagation mechanisms of floods during the peak phase. Therefore, using the third reservoir's inflow forecast during the peak flood season can improve forecast accuracy.

[0087] During the later stages of the water receding phase, the forecast will switch to the more stable inflow forecast for the second reservoir. Historical similarity scenarios, by referencing flow changes under similar past conditions, can provide a relatively stable reference for flow forecasting during the receding phase, as the receding process usually exhibits certain regularities, and the inflow forecast for the second reservoir can effectively simulate this process.

[0088] If the fusion model is determined not to be the optimal mode or the segmented mode, then the fusion model is determined to be the weighted average mode. According to the dynamic weights of the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir, the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir are weighted and averaged to obtain the final reservoir inflow forecast result.

[0089] Figure 2 These are schematic diagrams of modules for a hydrological forecasting system based on an adaptive correction chain, as shown in some embodiments of this specification. Figure 2 As shown, a hydrological forecasting system based on an adaptive correction chain may include a first forecasting module, a second forecasting module, a third forecasting module, and a forecast fusion module.

[0090] The first forecast module is used to acquire rainfall observation data and generate the inflow forecast result of the first reservoir based on the rainfall deviation source-dynamic correction chain and rainfall observation data;

[0091] The second forecast module is used to determine the multi-dimensional features of the current scene and generate the inflow forecast results of the second reservoir based on the historical similar sample matching-pattern application chain and the multi-dimensional features of the current scene.

[0092] The third forecast module is used to obtain underlying surface conditions, spatial distribution of rainfall and watershed status, and generate the inflow forecast results of the third reservoir based on the knowledge of runoff generation and confluence - threshold constraint chain and underlying surface conditions, spatial distribution of rainfall and watershed status.

[0093] The forecast fusion module is used to determine the dynamic weights of the inflow forecasts for the first, second, and third reservoirs, and to generate the final reservoir inflow forecast based on the dynamic weights and strategy selection tree of the inflow forecasts for the first, second, and third reservoirs.

[0094] The hydrological forecasting system based on adaptive correction chains can apply the aforementioned hydrological forecasting method based on adaptive correction chains, which will not be elaborated further here.

[0095] Finally, it should be understood that the embodiments described in this specification are merely illustrative of the principles of the embodiments described herein. Other variations may also fall within the scope of this specification. Therefore, alternative configurations of the embodiments described herein are intended to be illustrative rather than limiting, and should be considered consistent with the teachings of this specification. Accordingly, the embodiments described herein are not limited to those explicitly introduced and described herein.

Claims

1. A hydrological forecasting method based on adaptive correction chains, characterized in that, include: Acquire rainfall observation data, and generate the first reservoir inflow forecast result based on the rainfall deviation source-dynamic correction chain and rainfall observation data. This includes: aligning the observation data with the forecast grid using a spatiotemporal interpolation algorithm, calculating the areal rainfall error sequence, determining the error statistical characteristics, determining the correction parameters based on the error statistical characteristics, correcting the initial reservoir inflow forecast result based on the correction parameters, and generating the first reservoir inflow forecast result. The multi-dimensional features of the current scene are determined, and the inflow forecast results of the second reservoir are generated based on the historical similar sample matching-pattern application chain and the multi-dimensional features of the current scene. This includes: determining similar samples based on the multi-dimensional features of the current scene and obtaining the key hydrological feature values ​​of the similar samples. The underlying surface conditions, spatial distribution of rainfall, and watershed status are obtained. Based on the knowledge-threshold constraint chain of runoff generation and confluence, and the underlying surface conditions, spatial distribution of rainfall, and watershed status, the inflow forecast results of the third reservoir are generated. This includes generating the inflow forecast results of the third reservoir based on the optimal runoff generation parameters and the unit line of confluence using a third forecast model. The third forecast model embeds a physical constraint loss function. Determine the dynamic weights of the inflow forecast results for the first reservoir, the second reservoir, and the third reservoir; Based on the dynamic weights and strategy selection trees of the inflow forecasts for the first, second, and third reservoirs, the final reservoir inflow forecast results are generated. This includes: adjusting the basic weights of the inflow forecasts for the first, second, and third reservoirs based on the physical mechanisms of the current state, and determining the dynamic weights of the inflow forecasts for the first, second, and third reservoirs.

2. The hydrological forecasting method based on adaptive correction chain according to claim 1, characterized in that, Based on the rainfall deviation source-dynamic correction chain and rainfall observation data, the inflow forecast results for the first reservoir are generated, including: The first prediction model generates the initial reservoir inflow forecast.

3. The hydrological forecasting method based on adaptive correction chain according to claim 2, characterized in that, Based on the statistical characteristics of the error, the correction parameters are determined, including: Based on real-time weather data and a weather-error pattern rule base, the error pattern is determined. Obtain the dynamic and thermal parameters of the current scene; Based on error statistics, error patterns, and dynamic-thermal parameters of the current scenario, similar scenarios are identified. Based on similar scenarios, determine the correction parameters.

4. The hydrological forecasting method based on adaptive correction chain according to claim 2, characterized in that, The multi-dimensional features of the current scenario include rainfall characteristics, watershed status, and weather conditions; Based on historical similar sample matching and pattern application chains and multi-dimensional features of the current scenario, the inflow forecast results for the second reservoir are generated, including: The second prediction model generates the inflow forecast results for the second reservoir based on the key hydrological feature values ​​of similar samples.

5. The hydrological forecasting method based on adaptive correction chain according to claim 4, characterized in that, Based on runoff generation and confluence knowledge—threshold constraint chain, underlying surface conditions, spatial distribution of rainfall, and watershed status, the inflow forecast results for the third reservoir are generated, including: The optimal runoff generation parameters are determined by using real-time weather data and underlying surface conditions; Based on the spatial distribution of rainfall and the state of the watershed, a unit line of confluence is generated.

6. The hydrological forecasting method based on adaptive correction chain according to claim 5, characterized in that, The dynamic weights of the inflow forecasts for the first, second, and third reservoirs are determined, including: Based on the historical forecast errors of the rainfall deviation source tracing-dynamic correction chain, historical similar sample matching-regulation application chain, and runoff generation and confluence knowledge-threshold constraint chain, the basic weights of the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir are determined.

7. The hydrological forecasting method based on adaptive correction chain according to claim 6, characterized in that, Based on the current physical mechanism, the basic weights of the inflow forecasts for the first, second, and third reservoirs are adjusted to determine their dynamic weights, including: Based on the similarity of similar scenarios, a similarity enhancement factor is generated for the forecast results of the inflow of the first reservoir. Based on the similarity of similar samples, a similarity enhancement factor is generated for the forecast results of the inflow into the second reservoir. Based on the time interval of similar samples, a time decay factor is determined for generating the forecast results of the inflow into the second reservoir. Based on the physical mechanism of the current state, determine the fit of the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir. Based on the similarity enhancement factor of the inflow forecast results of the first reservoir, the similarity enhancement factor and time decay factor of the inflow forecast results of the second reservoir, and the fit of the inflow forecast results of the first, second and third reservoirs, the basic weights of the inflow forecast results of the first, second and third reservoirs are adjusted to determine the dynamic weights of the inflow forecast results of the first, second and third reservoirs.

8. The hydrological forecasting method based on adaptive correction chain according to any one of claims 1-7, characterized in that, Based on the dynamic weights and strategy selection trees of the inflow forecasts for the first, second, and third reservoirs, the final reservoir inflow forecast results are generated, including: Based on the dynamic weights of the inflow forecasts for the first, second, and third reservoirs, a fusion mode is determined, which may be an optimal mode, a segmented mode, or a weighted average mode. Based on the fusion model, the inflow forecast results of the first reservoir, the second reservoir, and the third reservoir are merged to generate the final reservoir inflow forecast result.

9. The hydrological forecasting method based on adaptive correction chain according to claim 8, characterized in that, Based on the dynamic weights of the inflow forecasts for the first, second, and third reservoirs, a fusion model is determined, including: If the dynamic weights of the inflow forecasts for the first, second, or third reservoirs are greater than the preset weight threshold, then the fusion mode is the optimal selection mode. If the dynamic weights of the inflow forecasts for the first, second, and third reservoirs are less than or equal to a preset weight threshold, then based on the river network humidity index and the spatial heterogeneity of rainfall, it is determined whether the fusion mode is a segmented mode.

10. A hydrological forecasting system based on an adaptive correction chain, characterized in that, The hydrological forecasting method based on adaptive correction chain as described in claim 1 includes: The first forecast module is used to acquire rainfall observation data and generate the inflow forecast result of the first reservoir based on the rainfall deviation source-dynamic correction chain and rainfall observation data; The second forecast module is used to determine the multi-dimensional features of the current scene and generate the inflow forecast results of the second reservoir based on the historical similar sample matching-pattern application chain and the multi-dimensional features of the current scene. The third forecast module is used to obtain underlying surface conditions, spatial distribution of rainfall and watershed status, and generate the inflow forecast results of the third reservoir based on the knowledge of runoff generation and confluence - threshold constraint chain and underlying surface conditions, spatial distribution of rainfall and watershed status. The forecast fusion module is used to determine the dynamic weights of the inflow forecasts for the first, second, and third reservoirs, and to generate the final reservoir inflow forecast based on the dynamic weights and strategy selection tree of the inflow forecasts for the first, second, and third reservoirs.

Images