A precipitation data fusion correction method suitable for observing a rare complex terrain area

Abstract

The application discloses a kind of precipitation data fusion correction methods suitable for observing scarce complex terrain area, comprising the following steps: step 1: multi-source data acquisition and preprocessing;Step 2: machine learning main body correction of monthly scale system deviation;Step 3: adaptive learning and dynamic compensation of daily scale residual sequence;Step 4: uncertainty quantification and fusion product generation;Step 5: inversion compensation and iterative optimization based on uncertainty;Step 6: product inspection and output.The method of the application realizes the high-precision fusion and dynamic correction of multi-source precipitation data in complex terrain area, significantly improves the temporal and spatial resolution and accuracy of precipitation data;By introducing nonlinear modeling and adaptive mechanism, the ability to describe short-term changes and extreme precipitation events is enhanced;At the same time, a closed-loop framework of uncertainty quantification and feedback optimization is established, which effectively improves the reliability of the results and the ability to support application decision-making.

Description

Technical Field

[0001] This invention belongs to the field of water resource management and disaster early warning technology, and in particular relates to a method for fusion and correction of precipitation data suitable for observation of areas with scarce and complex terrain. Background Technology

[0002] High-precision, high-spatiotemporal-resolution precipitation data serves as the core foundational data driving hydrological and meteorological research, water resource management, climate risk assessment, and disaster early warning; its accuracy is paramount. However, acquiring high-quality precipitation data that meets application requirements in complex terrain areas with dramatic topographic relief and sparse observation stations (such as high-altitude mountains and remote islands) has always faced long-term and systemic technical challenges. These challenges are rooted in three interrelated levels: data sources, methodological principles, and system framework.

[0003] At the data source level, all single-type products suffer from inherent and insurmountable deficiencies. While ground-based station observations are often considered the "benchmark" for assessment, their spatial representativeness is severely inadequate in complex terrain areas. Not only are these stations scarce, but they are also mostly located in easily accessible low-altitude valleys, completely failing to capture the spatial heterogeneity of precipitation caused by dramatic topographic changes. Furthermore, the "capture error" inherent in traditional rain gauges under severe weather conditions such as strong winds and freezing rain further weakens the absolute accuracy of their data. As an important supplement, satellite remote sensing and global reanalysis products, while providing spatially continuous precipitation fields, typically have a relatively coarse spatial resolution (greater than or equal to 10 km), making it difficult to characterize the finely modulated precipitation distribution patterns by micro-topography. More importantly, these products generally exhibit significant systematic biases in complex terrain areas (such as overall overestimation or underestimation), and the physical inversion or data assimilation algorithms they rely on have inherent limitations in their applicability in mountainous regions. While regional climate models can achieve relatively high resolution through dynamic downscaling, their simulation results are heavily dependent on the selection of physical parameterization schemes, introducing "structural errors" that are difficult to eliminate. At the same time, their high computational cost also hinders their widespread application in operational scenarios.

[0004] At the methodological level, existing technological approaches aimed at fusing and correcting multi-source data also face significant bottlenecks. Current mainstream methods can be broadly categorized into two types: physical statistical methods and purely data-driven methods. The former, such as interpolation or multiple linear regression based on topographic relationships, while possessing relatively clear physical interpretations, suffers from the core assumption of linearity or quasi-linearity, making it difficult to characterize the complex nonlinear coupling relationship between precipitation and multiple factors (such as topographic elevation, slope and aspect, atmospheric circulation, etc.). Furthermore, the correction effect rapidly diminishes under extreme terrain conditions. The latter, machine learning methods represented by random forests and neural networks, demonstrate advantages in capturing complex nonlinear patterns, but their current applications are mostly limited to "static single-scale correction" models. This approach typically uses monthly or annual average observations to train a fixed model to correct for long-term average biases. However, it suffers from two major drawbacks: First, the model cannot adaptively capture high-frequency signals. Once trained, its parameters are fixed and cannot be dynamically adjusted based on real-time or short-term weather system evolution, resulting in a severe deficiency in capturing crucial details of daily precipitation fluctuations, especially extreme precipitation events. Second, the method generally lacks a mechanism for recognizing and responding to uncertainty, usually only outputting a single "optimal estimate" without quantifying or providing a confidence range for the correction result itself. When the model is applied to extreme weather conditions or unfamiliar geographical environments not adequately covered by the training data, its output may exhibit severe bias and lack self-warning capabilities, posing hidden risks to subsequent applications.

[0005] At the system level, existing technical solutions generally lack a closed-loop intelligent framework from data input to high-quality product decision-making. Most solutions are "open-loop" unidirectional processing flows, namely "input data - execute correction - output product," which lack real-time evaluation of the output product quality, quantification of accompanying uncertainties, and dynamic feedback and compensation mechanisms based on the evaluation results. This open-loop structure makes the final data product a "black box" in terms of reliability, making it difficult for users to judge its credibility at a specific time and place, thus failing to provide solid support for high-confidence scientific decision-making and accurate risk warning.

[0006] In summary, current technologies have not yet provided a complete solution for generating precipitation data with high spatiotemporal resolution, high accuracy, adaptability to synoptic-scale changes, and self-assessment and reliability indication capabilities in complex terrain areas with scarce observations. This technological gap continues to hinder the deepening of research and the improvement of precise decision-making capabilities in the fields of hydrometeorology, climate change, and related resource management. Summary of the Invention

[0007] The purpose of this invention is to provide a method for fusing and correcting precipitation data suitable for observation of sparse and complex terrain areas, so as to solve the above-mentioned technical problems.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] This invention discloses a method for fusing and correcting precipitation data suitable for areas with scarce and complex topography. The method includes the following steps:

[0010] Step 1: Multi-source data acquisition and preprocessing: Acquire multi-source precipitation data for the study area, including: daily precipitation observation data from surface meteorological stations, daily precipitation data from the ERA5-Land reanalysis released by the European Centre for Medium-Range Weather Forecasts (ECMWF), and daily precipitation data simulated by the WRF regional climate model; the daily precipitation data simulated by the WRF regional climate model has a higher resolution than the daily precipitation data from the ERA5-Land reanalysis.

[0011] Daily precipitation data from the ERA5-Land reanalysis were physically constrained and downscaled using daily precipitation data simulated by the WRF regional climate model to generate daily precipitation data. Its core formula is the proportional downscaling method:

[0012]

[0013] In the formula, For grid points Downscaled daily precipitation; For grid points The original ERA5-Land daily precipitation; and Grid points WRF and ERA5-Land annual average precipitation;

[0014] Step 2: Machine Learning Correction of Monthly-Scale Systematic Bias: Based on ensemble learning algorithms, a monthly-scale precipitation correction model is constructed to eliminate systematic biases in daily-scale precipitation data; the specific process is as follows:

[0015] 1) Feature Construction: For each grid point, construct a feature vector containing geographical factors and climate circulation factors. ;

[0016] 2) Model training: The training target is the monthly average precipitation observed at ground stations. The monthly average precipitation data and eigenvectors at the corresponding grid points Use the input to train the RF regression model ;

[0017] 3) Generate a lunar-scale calibration field: Apply the trained model Predictions are made for all grid points across the entire region to generate monthly precipitation data after eliminating systematic biases. ;

[0018] 4) Main Body Correction: Monthly precipitation data is used to perform a proportional correction on daily precipitation data. The formula is:

[0019]

[0020] In the formula, For the corrected grid points Daily precipitation; For grid points Monthly precipitation; Daily precipitation data The monthly average for the corresponding month;

[0021] Step 3: Adaptive Learning and Dynamic Compensation of Daily-Scale Residual Sequences: By introducing signal decomposition techniques and time series deep learning, a dynamic model capable of adaptively learning and compensating for high-frequency residuals is constructed; the specific process includes the following steps:

[0022] Step 3-1: Extraction and construction of daily-scale residual sequences:

[0023] Define diurnal residual sequences at grid locations with ground observation stations. Its physical meaning is the estimated precipitation value after step 2 (monthly scale correction). Compared with actual observations at the station Deviation on a daily scale:

[0024]

[0025] In the formula, For site indexing, A collection of all available sites. Total number of days;

[0026] Step 3-2: Residual multi-scale decomposition based on empirical mode decomposition:

[0027] To separate physical processes at different time scales in the residual sequence, the Empirical Mode Decomposition (EMD) method is used. Decomposed into a series of intrinsic mode functions (IMFs) and a residual term:

[0028]

[0029] In the formula, Indicates the first The IMF components are arranged from highest to lowest frequency. Highest frequency; J represents the number of IMF components; The residual term represents the long-term trend or mean of the sequence;

[0030] EMD decomposition separates the original residuals into: high-frequency random components and synoptic-scale components. Low-frequency trend components ;

[0031] Step 3-3: LSTM residual prediction model enhanced by attention mechanism:

[0032] To dynamically predict weather-scale residuals for future moments Construct an LSTM residual prediction model that incorporates a self-attention mechanism;

[0033] Steps 3-4: Dynamic compensation and final daily precipitation data reconstruction:

[0034] Using the trained LSTM residual prediction model, denoted as For any future date Rolling predictions are performed using synoptic-scale residuals to obtain... By combining the low-frequency trend components, the complete prediction residual is reconstructed. The predicted daily residuals are dynamically superimposed onto the monthly correction field to ultimately generate high-resolution daily precipitation data with adaptive capabilities.

[0035]

[0036] In the formula, For the final generated grid points Daily precipitation; This represents a spatial interpolation function that will be used at station locations. Residuals predicted Interpolate to all high-resolution grid points superior;

[0037] Step 4: Uncertainty Quantification and Fusion Product Generation: The monthly structural uncertainty derived from the monthly scale correction model and the daily process uncertainty derived from the daily scale dynamic compensation are quantified respectively, and the total uncertainty is synthesized to finally generate a probabilistic precipitation product.

[0038] Step 5: Uncertainty-based inversion compensation and iterative optimization: Construct a feedback loop to optimize the results in the high-uncertainty region; specifically including the following processes:

[0039] 1) Threshold determination: Set an uncertainty threshold. When a certain grid point During the period Average uncertainty within When this occurs, the area is determined to be a high-uncertainty area, triggering a compensation mechanism;

[0040] 2) Variational inversion compensation: Constructing a cost function under the constraints of physical conservation and spatial smoothness. :

[0041]

[0042] In the formula, The variance represents the uncertainty. For smoothing constraint coefficients; For smoothness constraints;

[0043] By minimizing The optimized precipitation sequence for this region was obtained by solving the problem. Replace the original probabilistic precipitation products;

[0044] Step 6: Product Inspection and Output: Perform internal consistency checks on the final product and output the final precipitation dataset.

[0045] Furthermore, each IMF component in step 3 The following two conditions must be met to obtain the result through an iterative "screening" process:

[0046] 1) Throughout the entire data range, the number of extreme points must be equal to or at most differ by one from the number of zero-crossing points;

[0047] 2) At any given time, the mean of the upper envelope defined by the local maxima and the lower envelope defined by the local minima is zero.

[0048] Furthermore, the specific construction process of the LSTM residual prediction model in step 3 is as follows:

[0049] 1) Forward propagation of LSTM cells:

[0050] For time step Corresponding date The calculation of the LSTM unit is as follows:

[0051]

[0052] in, for The input feature vector at time step includes: the anomaly index of the previous circulation, the target grid point, and surrounding grid points. The local instability energy provided by short-time series and reanalysis data; It is the sigmoid activation function. This indicates element-wise multiplication;

[0053] 2) Self-attention mechanism:

[0054] To enhance the model's focus on key weather process signals, the output sequence of the LSTM layer... Apply a self-attention layer:

[0055] (6)

[0056] In the formula, ; The projection matrix is ​​learnable; This is the scaling factor.

[0057] Furthermore, the specific process for quantifying the monthly-scale structural uncertainty in step 4 is as follows:

[0058] Based on random forest ensemble, for each grid point Monthly scale forecast Its uncertainty stems from the composition of the forest. The degree of dispersion of the predicted values ​​of each decision tree;

[0059] 1) Prediction of a single tree:

[0060] make Indicates the first Each decision tree is used to process the input feature vector. Monthly precipitation forecast output;

[0061] 2) Integrated prediction and variance calculation:

[0062] The final prediction of a random forest is the mean of the predictions from all trees.

[0063] (8)

[0064] Its prediction variance, or structural uncertainty variance, is estimated by calculating the sample variance of all tree predictions:

[0065]

[0066] The specific process for quantifying the uncertainty of the diurnal process in step 4 is as follows:

[0067] Daily-scale residual prediction model The uncertainty in the process stems from the uncertainty of the model weights and the inherent randomness of weather systems. The Monte Carlo Dropout method is used to address this. Inferences are made based on the uncertainties of the process, and the specific process is as follows:

[0068] 1) Dropout regularization during training:

[0069] Dropout was enabled in the fully connected layers and other locations during the training of the LSTM-Attention network, with a dropout rate of [missing information]. ;

[0070] 2) Monte Carlo sampling during prediction:

[0071] During the prediction phase, the Dropout layer is not turned off; for the same input sequence, [the following steps are performed]. The forward propagation prediction; due to the randomness of Dropout, each forward propagation is equivalent to randomly sampling a sub-network from a set of models with shared parameters for prediction, thus obtaining... Slightly different daily-scale residual prediction values ;

[0072] 3) Calculation of process uncertainty variance:

[0073] this The dispersion of the prediction results represents the uncertainty of the model's prediction process for the residuals at that grid point on that day; its variance is calculated as follows:

[0074]

[0075] The final residual prediction value is taken from this. Mean of the samples: ;

[0076] The specific process of synthesizing the total uncertainty in step 4 to finally generate probabilistic precipitation products is as follows:

[0077] The final daily precipitation The uncertainty is contributed by both monthly-scale structural uncertainty and daily-scale process uncertainty; considering the operational relationship of formula (2), the first-order error propagation theory is used for approximate synthesis.

[0078] because ,and ,Will Considered as the primary source of uncertainty, its variance is and process uncertainty Perform synthesis; assuming the error sources are independent, the total uncertainty variance is approximately:

[0079]

[0080] The final probabilistic precipitation product is generated, assuming that the total error follows a mean of 0 and a variance of 0. The normal distribution; therefore, for lattice points ,date supply:

[0081] Optimal estimate: ;

[0082] Confidence interval: The 95% confidence interval is... ;

[0083] Standard deviation: As a direct measure of uncertainty.

[0084] Furthermore, step 6, which involves performing an internal consistency check on the final product and outputting the final precipitation dataset, specifically involves checking the spatiotemporal continuity and physical rationality of the final precipitation field; if all checks pass, the final precipitation dataset is output; if any part fails the check, the relevant information is fed back to step 5 or step 3 for limited iterative adjustments until the preset quality standards are met.

[0085] The beneficial effects of this invention are: the method described in this invention achieves high-precision fusion and dynamic correction of multi-source precipitation data in complex terrain areas, significantly improving the spatiotemporal resolution and accuracy of precipitation data; by introducing nonlinear modeling and adaptive mechanisms, it enhances the ability to characterize short-term changes and extreme precipitation events; at the same time, it constructs a closed-loop framework for uncertainty quantification and feedback optimization, effectively improving the reliability of results and the ability to support application decisions.

[0086] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0087] Figure 1 This is a flowchart of the method described in this invention;

[0088] Figure 2 This is a spatial distribution characteristic map of precipitation data in the study area before and after correction in Example 1. Detailed Implementation

[0089] This invention discloses a method for fusing and correcting precipitation data suitable for observation of areas with scarce and complex topography, such as... Figure 1 As shown, the method includes the following steps:

[0090] Step 1: Acquisition and preprocessing of multi-source data: Acquire multi-source precipitation data for the study area, including: daily precipitation observation data from ground meteorological stations, daily precipitation data from the ERA5-Land reanalysis released by the European Centre for Medium-Range Weather Forecasts (ECMWF) (usually at 10km resolution), and high-resolution (usually at 3km resolution) daily precipitation data simulated by the WRF regional climate model.

[0091] High-resolution precipitation data can depict more spatial details because its cell grid resolution is finer, better reflecting the topographic features of precipitation. High-resolution diurnal precipitation data were generated by physically constraining downscaling ERA5-Land precipitation data using high-resolution WRF model precipitation data. Its core formula is the proportional downscaling method:

[0092]

[0093] In the formula, For grid points Downscaled daily precipitation; For grid points The original ERA5-Land daily precipitation; and Grid points The WRF and ERA5-Land annual average precipitation.

[0094] This step unifies precipitation data to a high resolution (3km resolution) and incorporates higher-precision spatial features characterized by the WRF model.

[0095] Step 2: Machine Learning Correction of Monthly-Scale Systematic Bias: Based on ensemble learning algorithms such as Random Forest (RF), a monthly-scale precipitation correction model is constructed to eliminate systematic biases in daily-scale precipitation data; the specific process is as follows:

[0096] 1) Feature Construction: For each high-resolution grid point, a feature vector is constructed that includes geographical factors (longitude, latitude, elevation, slope, aspect) and climate circulation factors (such as convective available potential energy (CAPE) and total water vapor content (TCWV) from daily precipitation data from ERA5-Land reanalysis). .

[0097] 2) Model training: The monthly average observed precipitation at ground stations (calculated from daily precipitation observation data from ground meteorological stations) is used as the training target. The monthly average precipitation data and eigenvectors at the corresponding grid points Use the input to train the RF regression model .

[0098] 3) Generate a lunar-scale calibration field: Apply the trained model Predictions are made for all grid points across the entire region to generate monthly precipitation data after eliminating systematic biases. .

[0099] 4) Main Body Correction: Monthly precipitation data is used to perform a proportional correction on daily precipitation data. The formula is:

[0100]

[0101] In the formula, For the corrected grid points Daily precipitation; For grid points Monthly precipitation; Daily precipitation data The monthly average for the corresponding month.

[0102] Step 3: Adaptive Learning and Dynamic Compensation of Daily-Scale Residual Sequences: This step aims to address the core deficiency in Step 2—the insufficient ability of the monthly-scale correction model to capture daily-scale weather processes and extreme precipitation events. By introducing signal decomposition techniques and time-series deep learning, a dynamic model capable of adaptively learning and compensating for high-frequency residuals is constructed. The specific process includes the following steps:

[0103] Step 3-1: Extraction and construction of daily-scale residual sequences:

[0104] Define diurnal residual sequences at grid locations with ground observation stations. Its physical meaning is the estimated precipitation value after step 2 (monthly scale correction). Compared with actual observations at the station Deviation on a daily scale:

[0105]

[0106] In the formula, For site indexing, A collection of all available sites. This represents the total number of days. This residual sequence... It contains information that monthly-scale models fail to explain, which is dominated by diurnal weather processes (such as convective systems and frontal passages) and random noise.

[0107] Step 3-2: Residual multi-scale decomposition based on empirical mode decomposition:

[0108] To separate the physical processes at different time scales in the residual sequence, an adaptive signal processing method, Empirical Mode Decomposition (EMD), is employed. EMD can separate non-stationary and nonlinear residual sequences. Decomposed into a series of intrinsic mode functions (IMFs) and a residue:

[0109]

[0110] In the formula, Indicates the first The IMF components are arranged from highest to lowest frequency. (Highest frequency); J represents the number of IMF components; The residual term represents the long-term trend or mean of the sequence.

[0111] Each IMF component The following two conditions must be met to obtain the result through an iterative "sifting" process:

[0112] 1) Throughout the entire data range, the number of extreme points (maximum and minimum values) must be equal to or at most differ by one from the number of zero-crossing points.

[0113] 2) At any given time, the mean of the upper envelope defined by the local maxima and the lower envelope defined by the local minima is zero.

[0114] The original residuals were separated into the following by EMD decomposition:

[0115] 1) High-frequency random components ( ): Primarily caused by measurement noise and unresolved microphysical processes, it is usually not modeled.

[0116] 2) Weather-scale components ( , usually corresponds ): This is related to the lifecycle of predictable weather systems (such as low-pressure systems and shear lines) and is the core objective of modeling and compensation in this step.

[0117] 3) Low-frequency trend component ( : This may reflect intra-seasonal oscillations or climate change signals, and has been largely captured by monthly-scale models.

[0118] Step 3-3: LSTM residual prediction model enhanced by attention mechanism:

[0119] To dynamically predict weather-scale residuals for future moments A Long Short-Term Memory (LSTM) residual prediction model incorporating a self-attention mechanism is constructed. This model can simultaneously capture long-term dependencies in sequences and critical time periods that are crucial for prediction. The specific construction process of the model is as follows:

[0120] 1) Forward propagation of LSTM cells:

[0121] For time step (Corresponding date) The calculation of the LSTM unit is as follows:

[0122]

[0123] in, for The input feature vector at time step includes: anomaly indices of the previous circulation (such as standardized zonal wind indices), the target grid point, and surrounding grid points. Short-time series and reanalysis data provide local instability energy (CAPE), etc. It is the sigmoid activation function. This indicates element-wise multiplication.

[0124] 2) Self-attention mechanism:

[0125] To enhance the model's focus on key weather process signals, the output sequence of the LSTM layer... Apply a self-attention layer:

[0126] (6)

[0127] In the formula, ; The projection matrix is ​​learnable; This is the scaling factor.

[0128] Steps 3-4: Dynamic compensation and final daily precipitation data reconstruction:

[0129] Using a trained LSTM residual prediction model (denoted as...) (This can be done for any future date) Rolling predictions are performed using synoptic-scale residuals to obtain... Combined with low-frequency trend components (which can be derived from the residuals of historical residuals) (Smooth extrapolation or interpretation by a monthly scale model) reconstructs the complete prediction residuals. (In practice, high-frequency noise that cannot be modeled is usually ignored). Finally, the predicted diurnal residuals are dynamically superimposed onto the monthly-scale correction field to generate adaptive high-resolution diurnal precipitation data:

[0130]

[0131] In the formula, For the final generated grid points Daily precipitation; This represents a spatial interpolation function (such as distance-inverse weighted or Kriging interpolation), which will only be used at locations with stations. Residuals predicted Interpolate to all high-resolution grid points This step enables the transition from intelligent learning at a "point" level to spatial extension across a "surface".

[0132] Step 4: Uncertainty Quantification and Fusion Product Generation: The core objective of this step is to provide a reliability measure, i.e., uncertainty, for each grid point and time interval of the final generated diurnal precipitation data. Based on this, a hybrid uncertainty quantification framework is proposed to separately evaluate the structural uncertainty derived from the monthly-scale correction model and the process uncertainty derived from the diurnal-scale dynamic compensation, and then synthesize them; specifically, the following processes are included:

[0133] Step 4-1: Quantification of monthly-scale structural uncertainty (based on random forest ensemble):

[0134] Random forests are an ensemble model that naturally provides estimations of uncertainty. For each grid point... Monthly scale forecast Its uncertainty stems from the composition of the forest. The degree of dispersion of the predicted values ​​of each decision tree.

[0135] 1) Prediction of a single tree:

[0136] make Indicates the first Each decision tree is used to process the input feature vector. Monthly precipitation forecast output based on (geographical and climatic factors).

[0137] 2) Integrated prediction and variance calculation:

[0138] The final prediction of a random forest is the mean of the predictions from all trees.

[0139] (8)

[0140] Its prediction variance (structural uncertainty variance) can be estimated by calculating the sample variance of all tree predictions:

[0141]

[0142] variance This intuitively reflects the degree of disagreement within the model regarding the monthly precipitation estimate for that grid point. The greater the disagreement, the more uncertain the model's prediction for that point is based on existing data and features. This typically occurs in regions where the feature space is sparse or the relationship between features and precipitation is ambiguous.

[0143] Step 4-2: Quantification of uncertainties in daily-scale processes:

[0144] Daily-scale residual prediction model The uncertainty in the process stems from the uncertainty of the model weights and the inherent randomness of weather systems. The Monte Carlo Dropout (MC Dropout) method is used to address this. This method infers the uncertainty of the process, which is a practical approach for Bayesian approximate inference in deep learning. The specific process is as follows:

[0145] 1) Dropout regularization during training:

[0146] Dropout was enabled in the fully connected layers and other locations during the training of the LSTM-Attention network, with a dropout rate of [missing information]. This forces the network to avoid over-reliance on any particular neural connection, thus improving its generalization ability.

[0147] 2) Monte Carlo sampling during prediction:

[0148] During the prediction phase, the Dropout layer is not turned off. For the same input sequence (e.g., weather conditions from the previous 10 days), the following steps are performed: The forward propagation prediction. Due to the randomness of Dropout, each forward propagation is equivalent to randomly sampling a subnetwork from a large set of models with shared parameters for prediction, thus obtaining... Slightly different daily-scale residual prediction values .

[0149] 3) Calculation of process uncertainty variance:

[0150] this The dispersion of the prediction results represents the uncertainty of the model's prediction process for the residuals at that grid point on that day. Its variance is calculated as follows:

[0151]

[0152] Meanwhile, the final residual prediction value is taken from this. Mean of the samples: This is usually more stable and accurate than a single prediction.

[0153] Step 4-3: Synthesis of Total Uncertainty and Generation of Probabilistic Precipitation Products:

[0154] The final daily precipitation The uncertainty is contributed by both monthly-scale structural uncertainty and daily-scale process uncertainty. Considering the operational relationship of formula (2), the first-order error propagation theory is used for approximate synthesis.

[0155] because ,and ,Will It is considered the main source of uncertainty (its variance is...). ), and process uncertainty Perform synthesis. Assuming the error sources are independent, the total uncertainty variance can be approximated as:

[0156]

[0157] In the formula, the scaling factor The monthly-scale uncertainty variance is "distributed" to each specific day; if the precipitation on a certain day is much higher than the monthly average, the uncertainty introduced by the monthly-scale model on that day will also be amplified accordingly.

[0158] The final product generated is a probabilistic precipitation product. The output of this invention is no longer a single value, but a probability distribution. Typically, it can be assumed that the total error follows a distribution with a mean of 0 and a variance of... It follows a normal distribution. Therefore, for grid points... ,date supply:

[0159] Optimal estimate: ;

[0160] Confidence interval: For example, a 95% confidence interval is... ;

[0161] Standard deviation: As a direct measure of uncertainty.

[0162] This generates a dual-output product combining "data and uncertainty." Users can not only know "how much rainfall," but also "how reliable this estimate is," providing a crucial information dimension for subsequent risk-sensitive decisions (such as flood warnings and water resource allocation).

[0163] Step 5: Uncertainty-based inversion compensation and iterative optimization: Construct a feedback loop to optimize the results in the high-uncertainty region; specifically including the following processes:

[0164] 1) Threshold determination: Set an uncertainty threshold. When a certain grid point During the period Average uncertainty within If the region is deemed to be a high-uncertainty region, the result is deemed unreliable, triggering a compensation mechanism.

[0165] 2) Variational inversion compensation: A cost function is constructed based on physical conservation (such as regional water balance) and spatial smoothness constraints. :

[0166]

[0167] In the formula, The variance represents the uncertainty. For smoothing constraint coefficients; This is a smoothness constraint term.

[0168] By minimizing The optimized precipitation sequence for this region was obtained by solving the problem. Replace the original highly uncertain results, i.e. the probabilistic precipitation products generated in step 4.

[0169] Step 6: Product Inspection and Output: Perform internal consistency checks on the final product, including checking the spatiotemporal continuity of the final precipitation field (e.g., no outliers) and physical plausibility (e.g., covariance with temperature). If all checks pass, output the final high-resolution, probabilistic, adaptively corrected precipitation dataset; if any part fails a check, feed the relevant information back to Step 5 or Step 3 for finite-iterative fine-tuning until the preset quality standards are met.

[0170] Through the above steps, this invention realizes the whole-chain processing from raw data to high-quality intelligent products, forming a complete closed loop of "physical downscaling - machine learning subject correction - deep learning dynamic compensation - uncertainty quantification - feedback optimization", which significantly improves the ability to generate reliable precipitation data in complex terrain areas with scarce observations.

[0171] Example 1

[0172] This embodiment uses the Tibetan Plateau precipitation data gap basin as the study area and employs the above-mentioned method to fuse and correct the precipitation data of the study area. Figure 2 As shown, the spatial distribution characteristics and changes in annual average precipitation of precipitation data in the Tibetan Plateau watershed with missing precipitation data are compared before and after correction. Figure a shows the original precipitation data, which has a relatively smooth overall spatial distribution. High precipitation values ​​are mainly concentrated in the southeast of the region, exhibiting a clear single high-value core. The regional average precipitation is 905 mm, reflecting the original data's insufficient characterization of spatial heterogeneity under complex terrain. Figure b shows the corrected precipitation data, where the spatial distribution is significantly more refined. While high-value areas are still mainly located in the southeast, they exhibit stronger spatial dispersion and patchy characteristics. Simultaneously, the precipitation gradients in the central and western regions and high-altitude areas are clearer, better reflecting the modulating effect of topography and climate conditions on precipitation distribution. The regional average precipitation has decreased to 768 mm, indicating that the correction process, while weakening the systematic overestimation of the original data, has significantly improved the rationality and physical consistency of the precipitation spatial structure.

[0173] Finally, it should be noted that the above description is only used to illustrate the technical solution of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred arrangement, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for fusing and correcting precipitation data suitable for observation of sparse and complex terrain areas, characterized in that, The method includes the following steps: Step 1: Multi-source data acquisition and preprocessing: Acquire multi-source precipitation data for the study area, including: daily precipitation observation data from surface meteorological stations, daily precipitation data from the ERA5-Land reanalysis released by the European Centre for Medium-Range Weather Forecasts (ECMWF), and daily precipitation data simulated by the WRF regional climate model; the daily precipitation data simulated by the WRF regional climate model has a higher resolution than the daily precipitation data from the ERA5-Land reanalysis. Daily precipitation data from the ERA5-Land reanalysis were physically constrained and downscaled using daily precipitation data simulated by the WRF regional climate model to generate daily precipitation data. Its core formula is the proportional downscaling method: In the formula, For grid points Downscaled daily precipitation; For grid points The original ERA5-Land daily precipitation; and Grid points WRF and ERA5-Land annual average precipitation; Step 2: Machine Learning Correction of Monthly-Scale Systematic Bias: Based on ensemble learning algorithms, a monthly-scale precipitation correction model is constructed to eliminate systematic biases in daily-scale precipitation data; the specific process is as follows: 1) Feature Construction: For each grid point, construct a feature vector containing geographical factors and climate circulation factors. ; 2) Model training: The training target is the monthly average precipitation observed at ground stations. The monthly average precipitation data and eigenvectors at the corresponding grid points Use the input to train the RF regression model ; 3) Generate a lunar-scale calibration field: Apply the trained model Predictions are made for all grid points across the entire region to generate monthly precipitation data after eliminating systematic biases. ; 4) Main Body Correction: Monthly precipitation data is used to perform a proportional correction on daily precipitation data. The formula is: In the formula, For the corrected grid points Daily precipitation; For grid points Monthly precipitation; Daily precipitation data The monthly average for the corresponding month; Step 3: Adaptive Learning and Dynamic Compensation of Daily-Scale Residual Sequences: By introducing signal decomposition techniques and time series deep learning, a dynamic model capable of adaptively learning and compensating for high-frequency residuals is constructed; the specific process includes the following steps: Step 3-1: Extraction and construction of daily-scale residual sequences: Define diurnal residual sequences at grid locations with ground observation stations. Its physical meaning is the estimated precipitation value after step 2 (monthly scale correction). Compared with actual observations at the station Deviation on a daily scale: In the formula, For site indexing, A collection of all available sites. Total number of days; Step 3-2: Residual multi-scale decomposition based on empirical mode decomposition: To separate physical processes at different time scales in the residual sequence, the Empirical Mode Decomposition (EMD) method is used. Decomposed into a series of intrinsic mode functions (IMFs) and a residual term: In the formula, Indicates the first The IMF components are arranged from highest to lowest frequency. Highest frequency; J represents the number of IMF components; The residual term represents the long-term trend or mean of the sequence; The original residuals were separated into high-frequency random components and synoptic-scale components through EMD decomposition. Low-frequency trend components ; Step 3-3: LSTM residual prediction model enhanced by attention mechanism: To dynamically predict weather-scale residuals for future moments Construct an LSTM residual prediction model that incorporates a self-attention mechanism; Steps 3-4: Dynamic compensation and final daily precipitation data reconstruction: Using the trained LSTM residual prediction model, denoted as For any future date Rolling predictions are made using synoptic-scale residuals to obtain... By combining the low-frequency trend components, the complete prediction residual is reconstructed. The predicted daily residuals are dynamically superimposed onto the monthly correction field to ultimately generate high-resolution daily precipitation data with adaptive capabilities. In the formula, For the final generated grid points Daily precipitation; This represents a spatial interpolation function that will be used at station locations. The predicted residuals Interpolate to all high-resolution grid points superior; Step 4: Uncertainty Quantification and Fusion Product Generation: The monthly structural uncertainty derived from the monthly scale correction model and the daily process uncertainty derived from the daily scale dynamic compensation are quantified respectively, and the total uncertainty is synthesized to finally generate a probabilistic precipitation product. Step 5: Uncertainty-based inversion compensation and iterative optimization: Construct a feedback loop to optimize the results in the high-uncertainty region; specifically including the following processes: 1) Threshold determination: Set an uncertainty threshold. When a certain grid point During the period Average uncertainty within When this occurs, the area is determined to be a high-uncertainty area, triggering a compensation mechanism; 2) Variational inversion compensation: Constructing a cost function under the constraints of physical conservation and spatial smoothness. : In the formula, The variance represents uncertainty. For smoothing constraint coefficients; For smoothness constraints; By minimizing The optimized precipitation sequence for this region was obtained by solving the problem. Replace the original probabilistic precipitation products; Step 6: Product Inspection and Output: Perform internal consistency checks on the final product and output the final precipitation dataset.

2. The precipitation data fusion and correction method suitable for observation of sparse and complex terrain areas according to claim 1, characterized in that, Each IMF component in step 3 The following two conditions must be met to obtain the result through an iterative "screening" process: 1) The number of extreme points is equal to or at most differs from the number of zero-crossing points across the entire data range; 2) At any given time, the mean of the upper envelope defined by the local maxima and the lower envelope defined by the local minima is zero.

3. The precipitation data fusion and correction method suitable for observation of sparse and complex terrain areas according to claim 1, characterized in that, The specific construction process of the LSTM residual prediction model in step 3 is as follows: 1) Forward propagation of LSTM cells: For time step Corresponding date The calculation of the LSTM unit is as follows: in, for The input feature vector at time step includes: the anomaly index of the previous circulation, the target grid point, and surrounding grid points. The local instability energy provided by short-time series and reanalysis data; It is the sigmoid activation function. This indicates element-wise multiplication; 2) Self-attention mechanism: To enhance the model's focus on key weather process signals, the output sequence of the LSTM layer... Apply a self-attention layer: (6) In the formula, ; The projection matrix is ​​learnable; This is the scaling factor.

4. The precipitation data fusion and correction method suitable for observation of sparse and complex terrain areas according to claim 1, characterized in that, The specific process of quantifying the monthly-scale structural uncertainty in step 4 is as follows: Based on random forest ensemble, for each grid point Monthly scale forecast Its uncertainty stems from the composition of the forest. The degree of dispersion of the predicted values ​​of each decision tree; 1) Prediction of a single tree: make Indicates the first Each decision tree is used to process the input feature vector. Monthly precipitation forecast output; 2) Integrated prediction and variance calculation: The final prediction of a random forest is the mean of the predictions from all trees. (8) Its prediction variance, or structural uncertainty variance, is estimated by calculating the sample variance of all tree predictions: The specific process for quantifying the uncertainty of the diurnal process in step 4 is as follows: Daily-scale residual prediction model The uncertainty in the process stems from the uncertainty of the model weights and the inherent randomness of weather systems. The Monte Carlo Dropout method is used to address this. Inferences are made based on the uncertainties of the process, and the specific process is as follows: 1) Dropout regularization during training: Dropout was enabled in the fully connected layers and other locations during the training of the LSTM-Attention network, with a dropout rate of [missing information]. ; 2) Monte Carlo sampling during prediction: During the prediction phase, the Dropout layer is not turned off; for the same input sequence, [the following steps are performed]. The forward propagation prediction; due to the randomness of Dropout, each forward propagation is equivalent to randomly sampling a sub-network from a set of models with shared parameters for prediction, thus obtaining... Slightly different daily-scale residual prediction values ; 3) Calculation of process uncertainty variance: this The dispersion of the prediction results represents the uncertainty of the model's prediction process for the residuals at that grid point on that day; its variance is calculated as follows: The final residual prediction value is taken from this. Mean of the samples: ; The specific process of synthesizing the total uncertainty in step 4 to finally generate probabilistic precipitation products is as follows: The final daily precipitation The uncertainty is contributed by both monthly-scale structural uncertainty and daily-scale process uncertainty; considering the operational relationship of formula (2), the first-order error propagation theory is used for approximate synthesis. because ,and ,Will Considered as the primary source of uncertainty, its variance is and process uncertainty Perform synthesis; assuming the error sources are independent, the total uncertainty variance is approximately: The final probabilistic precipitation product is generated, assuming that the total error follows a mean of 0 and a variance of 0. The normal distribution; therefore, for lattice points ,date supply: Optimal estimate: ; Confidence interval: The 95% confidence interval is... ; Standard deviation: As a direct measure of uncertainty.

5. The precipitation data fusion and correction method suitable for observation of sparse and complex terrain areas according to claim 1, characterized in that, Step 6, which involves performing an internal consistency check on the final product and outputting the final precipitation dataset, specifically involves checking the spatiotemporal continuity and physical rationality of the final precipitation field. If all checks pass, the final precipitation dataset is output. If any part fails a check, the relevant information is fed back to step 5 or step 3 for limited iterative adjustments until the preset quality standards are met.

Images