Plateau multi-source rainfall data fusion method and device based on deep bidirectional sequential network

Abstract

The invention provides a plateau multi-source rainfall data fusion method and device based on a deep bidirectional sequential network. The method comprises the following steps: acquiring historical multi-source rainfall data; constructing and training a bidirectional time sequence attention multi-source fusion network model, and performing feature extraction and fusion calculation training; and performing precipitation estimation on the target area by using the bidirectional time sequence attention multi-source fusion network model, and outputting fused precipitation grid point data. Through multi-branch feature fusion, the network shows good nonlinear expression ability when processing a multi-scale rainfall field, the estimation precision of complex terrains and extreme rainfall events is effectively improved, a lightweight design is adopted, the network greatly reduces the calculation complexity while guaranteeing the multi-source rainfall fusion precision, and the calculation efficiency is improved. The method improves the flexibility of practical application, improves the fusion precision of multi-channel information such as radar reflectivity, satellite brightness temperature and ground rainfall, and greatly improves the precision and calculation efficiency of multi-source rainfall fusion.

Description

Technical Field

[0001] This application relates to the field of multi-source precipitation data fusion and prediction technology, and in particular to a method, apparatus, equipment, computer-readable medium and computer program product for high-altitude multi-source precipitation data fusion based on a deep bidirectional time series network. Background Technology

[0002] Multi-source precipitation data fusion is a core component in improving the accuracy of hydrological simulation and disaster early warning, significantly impacting drought and flood monitoring, water resource management, climate change response, and sustainable development. It is also a key method for assessing regional water cycle processes and extreme precipitation events. This fusion technology has been widely applied in flood forecasting, agricultural irrigation scheduling, urban flooding risk assessment, and hydropower station operation scheduling, as well as in research on extreme precipitation event detection, drought monitoring, and radar-satellite joint calibration. While traditional single-station observations offer high accuracy, they suffer from spatial sparsity and poor representativeness. With the development of remote sensing and numerical models, multi-source precipitation products such as satellite inversion, weather radar, and reanalysis data are gradually becoming the mainstream information sources.

[0003] Existing single-source precipitation data suffers from inconsistent spatiotemporal resolution, systematic biases, and random errors, leading to decreased accuracy in complex terrain areas and under extreme events. Furthermore, data heterogeneity caused by differences in the observation principles of different sensors may result in poor local performance of the fused data, ultimately affecting the reliability of hydrological simulations and disaster early warnings. Multi-source fusion, through joint inversion, Bayesian weighting, and deep learning, performs pixel-to-temporal joint correction on multi-sensor information, better capturing spatial details and evolution of precipitation, reducing spectral mixing and observational blind spots. Simultaneously, deep neural networks can automatically learn the multi-source error distribution characteristics, adapting to different climates and terrains, improving the generalization ability and spatiotemporal continuity of the fused product, thereby achieving higher accuracy, finer resolution, and more comprehensive precipitation data estimation.

[0004] Existing multi-source precipitation data fusion methods mainly include single-sensor estimation based on satellite infrared / microwave inversion, such as TMPA, CMORPH, and IMERG. These products utilize the scattering and radiation characteristics of cloud and rain particles at different wavelengths to generate areal rainfall through lookup tables or physical inversion algorithms. Due to their spatial continuity and wide coverage, they are widely used in flood monitoring and climate diagnostics. Another common approach is the ground-satellite joint bias correction method. This method uses probability matching, cumulative distribution function (CDF) matching, or Bayesian models to perform scale matching and systematic error correction between rain gauge observations and satellite products, achieving information transfer from station to grid point. This can effectively reduce the systematic bias of remote sensing products and improve local accuracy. Fusion frameworks based on numerical weather prediction (NWP) and radar assimilation have also received widespread attention. Variational or ensemble Kalman filtering incorporates radar reflectivity, satellite brightness temperature, and ground rainfall into the model background field, inverting the precipitation field under dynamic constraints, providing a higher spatiotemporal resolution initial field for short-duration heavy precipitation. In recent years, machine learning and deep learning technologies have made significant progress in multi-source precipitation fusion. By using convolutional neural networks, spatiotemporal sequence models, or attention mechanisms, these methods can automatically learn the complex nonlinear relationships between radar, satellite, ground, and terrain auxiliary variables, achieving pixel-to-temporal joint correction. This effectively captures local storm structures and extreme events, improving the spatial detail, temporal continuity, and adaptability to complex underlying surfaces and climate zones of the fusion products. They are suitable for high-resolution hydrological simulation and disaster early warning scenarios.

[0005] While existing multi-source precipitation data fusion methods have achieved some success, significant shortcomings remain. Products based on single-satellite inversion (such as TMPA and IMERG) rely on infrared / microwave brightness temperature. Although they offer wide coverage and rapid updates, they are significantly affected by sensor perspective, complex terrain, and snow cover, leading to systematic underestimation or overestimation. Furthermore, the spatiotemporal resolution is mismatched with the rain gauge scale, resulting in a sharp drop in accuracy in mountainous areas and during extreme events. Fusion methods based on ground-satellite bias correction, while reducing bias through CDF matching or Bayesian models, are highly dependent on station density and distribution. Correction signals are weak in sparse data areas, and thresholds and window parameters require manual setting, making them susceptible to experience-based influences and exhibiting poor cross-basin mobility. Radar-NWP assimilation frameworks, which assimilate multi-source observations through variational or ensemble Kalman filtering, can provide high-resolution precipitation fields. However, background field error setting, observation error covariance estimation, and microphysical scheme selection all introduce additional uncertainties, resulting in high computational load and long processing times, making it difficult to meet real-time operational needs. Moreover, the assimilation effect drops sharply in areas with radar coverage gaps or complex terrain. While the emerging machine learning and deep learning fusion technologies can automatically learn nonlinear mappings between radar, satellite, ground, and terrain auxiliary variables, enhancing spatial detail capture capabilities, they heavily rely on large-scale, high-quality training samples. This results in complex model structures, hyperparameter sensitivity, and a high susceptibility to overfitting. Furthermore, changes in climate zones, sensor replacements, or precipitation mechanisms can lead to decreased generalization performance, and insufficient interpretability hinders operational trust and deployment. Therefore, continuously optimizing algorithm robustness, reducing reliance on dense samples, and improving cross-basin and cross-sensor generalization capabilities remain key areas requiring breakthroughs in current multi-source precipitation fusion research.

[0006] In summary, existing methods often rely on single-sensor inversion or simple bias correction models, which are affected by sensor perspective, terrain shading, and snow cover, leading to systematic biases in precipitation estimation results in mountainous areas and during extreme events. Furthermore, traditional assimilation frameworks are computationally intensive, rely on human experience for parameters, and have poor real-time performance, making them unsuitable for high spatiotemporal resolution hydrological simulation and disaster early warning. Summary of the Invention

[0007] This application provides a method, apparatus, device, computer-readable medium, and computer program product for fusing multi-source precipitation data in high-altitude areas based on deep bidirectional temporal networks. It aims to solve the problems of insufficient accuracy, weak generalization ability, and low computational efficiency in existing technologies for fusing multi-source precipitation data. By introducing parallel computing and deep learning technologies, this invention aims to improve the accuracy, spatial detail, and computational efficiency of precipitation fusion, overcome the limitations of existing technologies, and provide a more flexible, robust, and efficient solution.

[0008] To achieve the above-mentioned technical effects, one aspect of this application provides a method for fusing plateau multi-source precipitation data based on deep bidirectional time-series networks, including: acquiring historical multi-source precipitation data;

[0009] Construct and train a bidirectional temporal attention multi-source fusion network model, and perform feature extraction and fusion computation training;

[0010] A bidirectional temporal attention multi-source fusion network model is used to estimate precipitation in the target area and output fused precipitation grid data.

[0011] According to a preferred embodiment of the present invention, the multi-source precipitation data to be fused further includes:

[0012] ERA5 reanalysis data, GLDAS land surface model data, and GPM satellite precipitation data, as well as topographic data, atmospheric temperature and humidity, air pressure, and wind speed data.

[0013] According to a preferred embodiment of the present invention, the bidirectional temporal attention multi-source fusion network model includes:

[0014] A bidirectional temporal convolutional neural network module is used to extract temporal and spatial features from precipitation sequences through bidirectional convolution;

[0015] The bidirectional long short-term memory network module is used to capture long-term evolutionary relationships based on the temporal features and output a bidirectional hidden state vector sequence.

[0016] According to a preferred embodiment of the present invention, the bidirectional temporal attention multi-source fusion network model further includes:

[0017] A cross-modal attention module is used to associate the temporal features with the original features and perform attention weight calibration;

[0018] The site embedding module is used to acquire precipitation data from meteorological stations, learn the unique representation vector of each meteorological station, and output spatial location correction information.

[0019] According to a preferred embodiment of the present invention, the bidirectional temporal attention multi-source fusion network model further includes a gated residual module, which is used to adaptively weight and fuse the temporal features and the original features through a gating mechanism.

[0020] According to a preferred embodiment of the present invention, the bidirectional temporal attention multi-source fusion network model further includes a global attention module. After the site embedding module outputs spatial location correction information, the global attention module obtains the bidirectional hidden state vector sequence output by the bidirectional long short-term memory network module and dynamically allocates the weights of the vectors in the sequence.

[0021] Another aspect of this application provides a plateau multi-source precipitation data fusion device based on a deep bidirectional time-series network, comprising:

[0022] Data acquisition module: used to acquire historical multi-source precipitation data;

[0023] Model building module: Used to build and train bidirectional temporal attention multi-source fusion network models, and to perform feature extraction and fusion calculation training;

[0024] The precipitation prediction module is used to estimate precipitation in the target area using a bidirectional temporal attention multi-source fusion network model and output fused precipitation grid data.

[0025] Another aspect of this application provides a plateau multi-source precipitation data fusion device based on a deep bidirectional time-series network, the device comprising:

[0026] At least one processor; and

[0027] A memory communicatively connected to the at least one processor; wherein,

[0028] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method as described above.

[0029] In another aspect, this application provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor to implement the above-described method.

[0030] In another aspect of this application, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described above.

[0031] The solution provided in this application has the following beneficial effects:

[0032] This invention employs a lightweight design, significantly reducing computational complexity while maintaining the accuracy of multi-source precipitation fusion, making it suitable for resource-constrained edge computing nodes and enhancing the flexibility of practical applications. The introduction of a gated residual network into the framework reduces the number of model parameters and computational costs while enhancing feature extraction capabilities, improving the fusion accuracy of multi-channel information such as radar reflectivity, satellite brightness temperature, and ground rainfall. Through multi-branch feature fusion, the network exhibits good nonlinear expression capabilities when processing multi-scale precipitation fields, effectively improving the estimation accuracy for complex terrain and extreme rainfall events. Furthermore, the application of normalization significantly enhances the model's stability, ensuring efficient convergence in different climate zones and sensor replacement scenarios. Overall, this invention significantly improves the accuracy and computational efficiency of multi-source precipitation fusion, providing solid technical support for real-time, high-resolution quantitative precipitation estimation and possessing broad application prospects. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0035] Figure 1 A flowchart illustrating a plateau multi-source precipitation data fusion method based on a deep bidirectional temporal network, provided as an embodiment of this application;

[0036] Figure 2 A diagram illustrating a bidirectional temporal attention multi-source fusion network model architecture is provided in one embodiment of this application.

[0037] Figure 3 A comparison chart of performance indicators of the method provided in one embodiment of this application and other precipitation fusion prediction models;

[0038] Figure 4 A comparison chart of monitoring capability indicators of the method provided in one embodiment of this application and other precipitation fusion prediction models;

[0039] Figure 5 A density scatter plot comparing the predicted and actual values ​​of the method provided in one embodiment of this application with those of other precipitation fusion prediction models;

[0040] Figure 6A comparison diagram of spatial distribution predictions of precipitation in Qinghai Province for a certain period of time using the method provided in an embodiment of this application and other precipitation fusion prediction models;

[0041] Figure 7 A schematic diagram of a plateau multi-source precipitation data fusion device based on a deep bidirectional time-series network is provided as an embodiment of this application;

[0042] Figure 8 This is a schematic diagram of the structure of a device suitable for implementing the solutions in the embodiments of this application.

[0043] The same or similar reference numerals in the accompanying drawings represent the same or similar parts. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0045] In a typical configuration of this application, the terminal and the service network devices each include one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0046] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0047] Computer-readable media include permanent and non-permanent, removable and non-removable media, which can store information by any method or technology. Information can be computer program instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, read-only optical disc (CD-ROM), digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0048] In practical scenarios, the execution entity of this method can be a user device, or a device formed by integrating a user device and a network device through a network, or it can be an application running on the aforementioned devices. The user device includes, but is not limited to, various terminal devices such as computers, mobile phones, tablets, smartwatches, and wristbands. The network device includes, but is not limited to, network hosts, single network servers, multiple network server sets, or cloud computing-based computer sets, and can be used to implement some processing functions when setting an alarm clock. Here, the cloud consists of a large number of hosts or network servers based on cloud computing. Cloud computing is a type of distributed computing, consisting of a virtual computer composed of a group of loosely coupled computer sets.

[0049] Figure 1 A flowchart illustrating a plateau multi-source precipitation data fusion method based on a deep bidirectional temporal network, as provided in one embodiment of this application, is shown below. Figure 1 As shown, the method includes at least the following processing steps:

[0050] Step S101: Obtain historical multi-source precipitation data.

[0051] Specifically, historical multi-source precipitation data and auxiliary data (topography, temperature, humidity, wind speed, atmospheric pressure, etc.) were downloaded using ERA5, GLDAS, and GPM, along with ground observation station data (meteorological variables and station coordinates: longitude, latitude, and altitude). The multi-source precipitation data and auxiliary data underwent preprocessing, including missing values, outliers, spatiotemporal resampling, and spatiotemporal alignment, followed by normalization. The final result was a dataset with a temporal resolution of three hours and a spatial resolution of 0.01. o ×0.01 o Precipitation dataset.

[0052] Step S102: Construct and train a bidirectional temporal attention multi-source fusion network model, and perform feature extraction and fusion calculation training.

[0053] Specifically, the preprocessed dataset is divided into a training set and a test set in a 7:3 ratio, with 70% of the data used to train the model and 30% used to test the model's performance.

[0054] Figure 2 A diagram of a bidirectional temporal attention multi-source fusion network model architecture is provided in one embodiment of this application, such as... Figure 2As shown, this embodiment proposes a method for constructing a plateau multi-source precipitation data collaborative inversion model (BTLA) based on deep bidirectional temporal network fusion, aiming to generate high-quality, high-resolution 3-hourly, 1-kilometer precipitation gridded data. By combining reanalysis data and ground observation station data, high-precision hourly precipitation estimation under complex terrain conditions is achieved. This framework can be represented as:

[0055]

[0056] Where, x E ,x GL and x GP Let f(.) represent the original precipitation datasets from the ERA5, GLDAS, and GPM datasets, respectively, and let f(.) denote the fusion network. Environmental variables are used as auxiliary inputs to support the fusion process and improve prediction accuracy, including longitude (lon), latitude (lat), DEM, surface atmospheric pressure (sp), wind speed (wd), specific humidity (sh), and 2m-temperature (tp). Furthermore, physical constraints are introduced. This can be used to improve the spatiotemporal correlation of precipitation, specifically as follows:

[0057]

[0058] in, , and These represent mass conservation constraints, thermodynamic constraints, and boundary constraints, respectively. This represents the predicted precipitation for the i-th sample. The value represents the actual precipitation of the i-th sample; N represents the number of samples in the batch; and S represents the length of the input time series. This represents the available precipitation for the i-th sample at time t; Indicates the water vapor convergence enhancement coefficient; Indicates the extreme precipitation threshold; This represents the temperature of the i-th sample; This represents the thermodynamic maximum precipitation rate, and the specific formula can be found in the Clausius-Clapeyron equation:

[0059]

[0060] in, This represents the saturated water vapor pressure at temperature T. Indicates the reference vapor pressure; a and b represent the coefficients of the Magnus formula; Indicates surface air pressure. Indicates the thickness of the troposphere; This indicates the density of water.

[0061] The BTLA model is designed to balance computational complexity and fusion accuracy, and consists of six main parts: Bidirectional Temporal Convolutional Neural Network (BiTCN) module, Bidirectional Long Short-Term Memory Network (BiLSTM) module, Cross-Modal Attention module, Site Embedding module, Gated Residual module, and Global Attention module.

[0062] First, the BiTCN layer extracts local and short-term dependency features from the precipitation sequence through bidirectional convolution. Then, the BiLSTM layer further captures long-term evolutionary relationships to achieve global modeling in the temporal dimension. Next, the model introduces cross-modal attention to dynamically weight the importance of different time steps and data sources, highlighting key feature information and enhancing the collaborative representation of multi-source data. Simultaneously, a gated residual structure with site embedding enables personalized calibration of precipitation patterns under different terrain and climatic conditions on the plateau, improving the model's adaptability to spatial heterogeneity. Finally, the global attention module, as the dynamic weight allocation part for key features within the temporal sequence, provides a stable and globally aware temporal context for the subsequent fully connected regression head.

[0063] BTLA uses ground-based observation data as a benchmark to guide the fusion calibration of satellite and reanalysis products, effectively improving the consistency and accuracy of precipitation estimation. The entire network framework is designed to achieve efficient estimation of multi-source precipitation data fusion, demonstrating significant potential for practical applications.

[0064] The Bidirectional Temporal Convolutional Neural Network (BiTCN) module serves as the input port of the framework, responsible for preliminary feature extraction from the input multi-source precipitation data and auxiliary data. It employs a bidirectional structure to extract information from both forward and backward directions. Through a series of convolution, pooling, and activation operations, the convolutional kernels are moved simultaneously in space and time to obtain features of different temporal and spatial information in the sequence data, effectively capturing local and global features in the sequence and improving the model's expressive power.

[0065] BiTCN emphasizes the effective extraction of spatiotemporal features both spatially and temporally, ensuring that the subsequent BiLSTM receives high-quality features. Furthermore, adding adaptive convolutions to the front and rear outputs of BiTCN to effectively fuse the features from both directions is crucial for subsequent feature learning. Features from both directions are fused using the adaptive convolution `Adaptive_Conv`.

[0066]

[0067]

[0068] in, It is a TCN bidirectional front-to-back fusion feature; This represents the hidden state of layer l at time t; This represents the weights of the convolutional kernel in layer l; Indicates the bias of the l-th layer; Indicates the kernel size; This represents the hidden state of layer l-1 at time t; This represents the hidden state of layer l-1 at time t-k+1; and These represent the weight matrix and the bias term, respectively. and These represent the forward feature layer and the backward feature layer, respectively.

[0069] The BiLSTM module, acting as a long-range dependency catcher in the network, is responsible for further modeling the cyclical evolution of the 24-hour seasonal time window on top of the "local-multi-scale" features provided by BiTCN. The module employs a two-layer bidirectional LSTM with 256 hidden dimensions. The output, after being normalized by LayerNorm, is concatenated with the attention output to form a composite temporal representation that simultaneously incorporates both "cyclical-attention" mechanisms. This structure effectively mitigates the gradient vanishing and memory decay problems that occur during long-sequence training by using 0.5 norm pruning and Dropout=0.3 regularization at the gradient level, providing a robust long-period context vector for the final precipitation estimation.

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] in, , , These represent the forget gate, input gate, and output gate, respectively. This represents the candidate cell state of the neuron at time t; Represents a memory unit; Indicates a hidden unit; , [x,y] represents the weights and biases; [x,y] represents the connection between two vectors. This represents the hidden layer of the backward LSTM; T represents the total length of the sequence; T-t+1 represents the reverse index of the time series.

[0079] The cross-modal attention module is used to associate the temporal features with the original features, calibrate the attention weights, and align the local convolutional features extracted by BiTCN with the original input sequence across modalities. This module uses the original sequence as the query and the BiTCN output as the key / value pair. It adaptively calibrates the weights at each time step through a scaled dot product attention mechanism, thereby suppressing noise channels and enhancing precipitation-sensitive physical variables (such as specific humidity and temperature-dew point difference). To prevent gradient anomalies, the attention score is strictly truncated to the [-50, 50] interval before softmax, and LayerNorm is introduced to pre-normalize the query / key, significantly reducing the risk of numerical explosion caused by extreme weather events (such as rainstorm peaks). This design enables the network to dynamically fuse "local-global" dual-view features without changing the convolutional skeleton, providing a more discriminative temporal representation for subsequent gating residuals.

[0080]

[0081] Where A represents the weight, This represents the dimension of the i-th query vector; This represents the i-th query matrix; The transpose of the i-th key matrix; Let represent the i-th value matrix; Softmax represents the normalization function.

[0082] The site embedding module is used to acquire precipitation data from meteorological stations and learn a low-dimensional, scalable, unique representation vector for each meteorological station. Combined with the spatial features extracted by BiTCN, the output precipitation intensity is subjected to location-related affine correction, and the output spatial location correction information is generated.

[0083] The site embedding module employs an nn.Embedding lookup table structure, outputting a 32-dimensional latent vector. Based on this, a dual-path lightweight sub-network is introduced to generate scale and shift, both compressed to [-1,1] using Tanh compression and then linearly mapped to the intervals [0.5,1.5] and [-0.5,0.5], respectively, achieving a gentle adjustment to the final regression value. This design, while maintaining global model parameter sharing, allows for some deviations between sites at different altitudes and climate zones, significantly reducing systematic bias (RB) caused by geographical location differences. Embedding weights are updated end-to-end with the main network, and prior information can be injected during the inference phase using only the site ID.

[0084] The gated residual module is responsible for performing adaptive weighted fusion between the deep features of BiTCN and the original input to solve the feature redundancy problem caused by the traditional equal-weighted addition of residuals. The module first concatenates the BiTCN output, the original sequence, and the site embedding vector along the channel dimension. Through two layers of linear mapping, a sigmoid function generates element-wise gate coefficients α∈[0,1]. Then, α is used to perform soft selection on the BiTCN features and the projected original input, achieving a proportional fusion of "convolution-original" information. To avoid gate saturation, the input vectors are standardized using LayerNorm before concatenation to ensure that the gradient maintains unit variance throughout the path. This mechanism can automatically reduce the dependence on the convolution path in extreme weather scenarios (such as prolonged drought followed by sudden flooding), retaining more original physical quantities, thereby improving the model's robustness and interpretability to abrupt events.

[0085] Finally, the global attention module, acting as the dynamic weight allocation part for key features within the temporal sequence, is responsible for mining global dependencies on the bidirectional hidden state sequence output by the BiLSTM, further aligning the time indices corresponding to extreme precipitation peaks. After the site embedding module outputs spatial location correction information, the global attention module obtains the bidirectional hidden state vector sequence output by the bidirectional long short-term memory network module and dynamically allocates the weights of the vectors in the sequence. Scaled dot-product attention is computed in parallel using four attention heads, allowing direct interaction between the hidden states at any two time points. This overcomes the local Markov assumption of LSTM, explicitly modeling cross-day causal chains such as "early high humidity - later strong convection." A 0.2 dropout is applied to the attention weights after softmax to prevent overfitting to individual heavy rainfall moments. The output is added to the LSTM hidden residuals before LayerNorm is applied to ensure that the network maintains zero mean and unit variance during inter-layer propagation, providing a stable and globally aware temporal context for subsequent fully connected regression heads.

[0086] The output of the global attention module is connected to a fully connected layer, which serves as the network's precipitation intensity regression prediction layer. This layer is responsible for mapping the fused 2048-dimensional time-site joint features to 1mm-level precipitation estimates. The module employs a three-layer feedforward structure (256→128→1), with each layer equipped with LayerNorm+GELU+Dropout to suppress overfitting while maintaining nonlinear expression. The input vector contains both global statistical information and retains local extrema and geographical priors, effectively improving the fitting ability for both 0.1mm micro-precipitation and extreme precipitation exceeding 50mm. The final linear output layer is connected to a site-specific scale-shift affine, and ReLU pruning ensures physical nonnegativity.

[0087] The above-mentioned bidirectional temporal attention multi-source fusion network model is used to train on the training set and the best results are verified through the test set.

[0088] Step S103: Use a bidirectional temporal attention multi-source fusion network model to estimate precipitation in the target area and output the fused precipitation grid data.

[0089] Specifically, taking Qinghai Province as an example, precipitation prediction and estimation were carried out for the entire region of Qinghai Province, and finally a precipitation dataset for Qinghai Province with a time resolution of three hours and a spatial resolution of 0.01x0.01o was generated.

[0090] Figure 3 A performance comparison chart of the method provided in one embodiment of this application and other precipitation fusion prediction models is shown, such as... Figure 3 As shown, this paper presents the performance of several existing methods (MLP, BPNN, CNN_LSTM, TCN, GNN, Transformer, and XGBoost), raw precipitation data (ERA5, GLDAS, and GPM), and the method proposed in this embodiment in precipitation fusion. CC (Correlation coefficient), KGE (Kling-Gupta Efficiency), RMSE (Root mean square error), and RB (Relative bias) are the most important indicators for evaluating the performance of the fusion model. Higher CC and KGE values ​​are better, while lower RMSE and RB values ​​are better. Compared with these seven common fusion methods, the method proposed in this patent achieves the highest performance across all four evaluation indicators.

[0091] Figure 4 A comparison chart of monitoring capability indicators of the method provided in one embodiment of this application and other precipitation fusion prediction models is shown, such as... Figure 4 As shown, Probability of Detection (POD), False Alarm Rate (FAR), and Critical Success Index (CSI) are commonly used indicators for evaluating the performance of meteorological event monitoring or forecasting models. They are calculated based on contingency tables (also known as confusion matrices) and are used to quantify the detection effect. The higher the POD value, the stronger the model's ability to capture real events; the lower the FAR value, the lower the model's false alarm rate; the CSI value ranges from 0 to 1, and the higher the value, the better the detection effect. It can be seen that the method in this embodiment has the best fusion effect.

[0092] Figure 5 A density scatter plot comparing the predicted and actual values ​​of the method provided in one embodiment of this application and other precipitation fusion prediction models, as shown in the figure. Figure 5As shown, the closer the scatter distribution is to the fitted line, the stronger the consistency between the predicted value and the observed data. It can be seen that the predicted value of the method in this embodiment is closest to the true value.

[0093] Figure 6 A comparison diagram of the spatial distribution prediction of precipitation in Qinghai Province for a certain period of time, provided by the method of this application and other precipitation fusion prediction models, is shown below. Figure 6 As shown, the data fused by the method proposed in this embodiment is closer to the actual measured values. In the complex plateau terrain of Qinghai Province, where precipitation is uneven and monitoring stations are sparsely distributed, this method can improve the model's generalization ability and stability.

[0094] The solution provided in this embodiment, through multi-branch feature fusion, demonstrates good nonlinear expression ability when processing multi-scale precipitation fields, effectively improving the estimation accuracy for complex terrain and extreme rainfall events. By adopting a lightweight design, the network significantly reduces computational complexity while ensuring the accuracy of multi-source precipitation fusion, enhancing the flexibility of practical applications and improving the fusion accuracy of multi-channel information such as radar reflectivity, satellite brightness temperature, and ground rainfall, thus greatly improving the accuracy and computational efficiency of multi-source precipitation fusion.

[0095] Figure 7 A schematic diagram of a plateau multi-source precipitation data fusion device based on a deep bidirectional time-series network is provided as an embodiment of this application, as shown below. Figure 7 As shown, the device includes:

[0096] Data acquisition module 11: Used to acquire multi-source precipitation data to be fused;

[0097] Model building module 22: Used to build and train a bidirectional temporal attention multi-source fusion network model, and to perform feature extraction and fusion calculation training;

[0098] The precipitation prediction module 33 is used to input the multi-source precipitation data to be fused into a bidirectional temporal attention multi-source fusion network model and output the fused precipitation grid data. The above device can execute the plateau multi-source precipitation data fusion method based on a deep bidirectional temporal network as described in the preceding embodiments, wherein...

[0099] The data acquisition module 11 executes step S101, the model building module 22 executes step S102, and the precipitation prediction module 33 executes step S103.

[0100] Based on the same inventive concept, this application also provides an electronic device. The method corresponding to the electronic device can be the plateau multi-source precipitation data fusion method based on deep bidirectional time-series networks in the foregoing embodiments, and its problem-solving principle is similar to that method. The electronic device provided in this application includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the methods and / or technical solutions of the foregoing embodiments of this application.

[0101] The electronic device can be a user device, or a device formed by integrating user devices and network devices through a network, or it can be an application running on the aforementioned devices. The user device includes, but is not limited to, various terminal devices such as computers, mobile phones, tablets, smartwatches, and wristbands. The network device includes, but is not limited to, network hosts, single network servers, multiple network server sets, or cloud computing-based computer sets, and can be used to implement some processing functions when setting an alarm clock. Here, the cloud consists of a large number of hosts or network servers based on cloud computing. Cloud computing is a type of distributed computing, consisting of a virtual computer composed of a group of loosely coupled computer sets.

[0102] Figure 8 The diagram illustrates the structure of an apparatus suitable for implementing the methods and / or technical solutions in the embodiments of this application. The apparatus 1200 includes a Central Processing Unit (CPU) 1201, which can perform various appropriate actions and processes based on a program stored in a Read Only Memory (ROM) 1202 or a program loaded from a storage portion 1208 into a Random Access Memory (RAM) 1203. The RAM 1203 also stores various programs and data required for system operation. The CPU 1201, ROM 1202, and RAM 1203 are interconnected via a bus 1204. An Input / Output (I / O) interface 1205 is also connected to the bus 1204.

[0103] The following components are connected to I / O interface 1205: an input section 1206 including a keyboard, mouse, touchscreen, microphone, infrared sensor, etc.; an output section 1207 including a cathode ray tube (CRT), liquid crystal display (LCD), LED display, OLED display, etc., and speakers, etc.; a storage section 1208 including one or more computer-readable media such as hard disk, optical disk, magnetic disk, semiconductor memory, etc.; and a communication section 1209 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 1209 performs communication processing via a network such as the Internet.

[0104] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by the central processing unit (CPU) 1201, it performs the functions defined in the methods of this application.

[0105] Another embodiment of this application provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application described above.

[0106] Specifically, this embodiment may employ any combination of one or more computer-readable media. A computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium may be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0107] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0108] The program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0109] Another embodiment of this application provides a computer program product, including a computer program that, when executed by a processor, implements a method for fusing multi-source precipitation data from plateau regions based on a deep bidirectional time-series network.

[0110] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0111] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0113] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or page components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0114] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0115] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units.

[0116] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

[0118] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a device claim may also be implemented by a single unit or device through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.

Claims

1. A method for fusing multi-source precipitation data from plateau regions based on deep bidirectional temporal networks, characterized in that, include: Obtain historical multi-source precipitation data; Construct and train a bidirectional temporal attention multi-source fusion network model, and perform feature extraction and fusion computation training; A bidirectional temporal attention multi-source fusion network model is used to estimate precipitation in the target area and output fused precipitation grid data.

2. The plateau multi-source precipitation data fusion method based on deep bidirectional temporal network according to claim 1, characterized in that, The acquisition of historical multi-source precipitation data further includes: ERA5 reanalysis data, GLDAS land surface model data, and GPM satellite precipitation data, as well as topographic data, atmospheric temperature and humidity, air pressure, and wind speed data.

3. The plateau multi-source precipitation data fusion method based on deep bidirectional temporal network according to claim 1, characterized in that, The bidirectional temporal attention multi-source fusion network model includes: A bidirectional temporal convolutional neural network module is used to extract temporal and spatial features from precipitation sequences through bidirectional convolution; The bidirectional long short-term memory network module is used to capture long-term evolutionary relationships based on the temporal features and output a bidirectional hidden state vector sequence.

4. The plateau multi-source precipitation data fusion method based on deep bidirectional temporal network according to claim 3, characterized in that, The bidirectional temporal attention multi-source fusion network model also includes: A cross-modal attention module is used to associate the temporal features with the original features and perform attention weight calibration; The site embedding module is used to acquire precipitation data from meteorological stations, learn the unique representation vector of each meteorological station, and output spatial location correction information.

5. The plateau multi-source precipitation data fusion method based on deep bidirectional temporal network according to claim 4, characterized in that, The bidirectional temporal attention multi-source fusion network model also includes a gated residual module, which is used to adaptively weight and fuse the temporal features and the original features through a gating mechanism.

6. The plateau multi-source precipitation data fusion method based on deep bidirectional temporal network according to claim 4, characterized in that, The bidirectional temporal attention multi-source fusion network model also includes a global attention module. After the site embedding module outputs spatial location correction information, the global attention module obtains the bidirectional hidden state vector sequence output by the bidirectional long short-term memory network module and dynamically allocates the weights of the vectors in the sequence.

7. A plateau multi-source precipitation data fusion device based on a deep bidirectional time-series network, characterized in that, include: Data acquisition module: used to acquire historical multi-source precipitation data; Model building module: Used to build and train bidirectional temporal attention multi-source fusion network models, and to perform feature extraction and fusion calculation training; The precipitation prediction module is used to estimate precipitation in the target area using a bidirectional temporal attention multi-source fusion network model and output fused precipitation grid data.

8. A plateau multi-source precipitation data fusion device based on a deep bidirectional time-series network, characterized in that, The device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 6.

9. A computer-readable medium having computer program instructions stored thereon, characterized in that, The computer program instructions can be executed by a processor to implement the method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images