Three-dimensional variational data assimilation method based on flow dependence model

Abstract

The invention provides a three-dimensional variation data assimilation method based on a flow dependence model, and the method comprises the following steps: S1, carrying out the preprocessing of multi-source data and the setting of a region-element adaptive time window: carrying out the preprocessing of multi-source meteorological data, and dynamically setting a time window [tau] based on the characteristics of meteorological elements and the characteristics of a region; s2, constructing and training a flow dependency model: constructing a deep learning flow dependency model of a CNN-LSTM hybrid architecture; and S3, constructing a cost function to carry out gradient solution. According to the three-dimensional variation data assimilation method based on the flow dependence model, a deep learning flow dependence model is introduced, a meteorological element association model in a small-scale time dimension is accurately constructed through a high-precision multi-feature extraction capability, traditional static prior information is replaced, and the new three-dimensional variation method has dynamic time adaptability; time windows are dynamically set for different regions and different meteorological elements for rapid model training, and the problem of adaptability of fixed time windows in a traditional method is solved.

Description

Technical Field

[0001] This invention relates to the application of deep learning in weather forecasting, and more particularly to a three-dimensional variational data assimilation method based on a flow-dependent model. Background Technology

[0002] Three-dimensional variational techniques are one of the core methods for meteorological data assimilation. By constructing an objective function and solving for the optimal solution, they achieve the fusion of multi-source observation data and the background field of numerical models, providing an initial field for numerical forecasting. Currently, deep learning is widely used to improve forecast accuracy and expand forecast dimensions. For example, CNNs, as fast recognition models, can efficiently identify features such as precipitation and cloud formations.

[0003] Traditional three-dimensional variational techniques are widely used in meteorological operations due to their moderate computational cost and high operational efficiency. However, they have key drawbacks: prior information is static and fixed, and cannot adapt to changes over time; at the same time, traditional methods struggle to accurately capture the correlations between meteorological elements on a small-scale time dimension, and they do not dynamically adapt time windows for different regions and meteorological elements. This makes it difficult for static prior information to meet the forecasting requirements of high-resolution numerical models, thus affecting assimilation accuracy and forecast performance. While four-dimensional variational techniques can consider the evolution of the time dimension, they rely on complex adjoint models, resulting in extremely high computational costs. Furthermore, they have stringent requirements for the temporal continuity of observational data, limiting their operational applications.

[0004] Therefore, it is necessary to combine deep learning to provide a three-dimensional variational data assimilation method based on a flow-dependent model to solve the above-mentioned technical problems. Summary of the Invention

[0005] This invention provides a three-dimensional variational data assimilation method based on a flow-dependent model, which solves the problems of statically fixed prior information, insufficient capture of small-scale time-dimensional relationships, and poor time window adaptability in traditional three-dimensional variational techniques, while avoiding the high computational cost of four-dimensional variational methods.

[0006] To address the aforementioned technical problems, this invention provides a three-dimensional variational data assimilation method based on a flow-dependent model, comprising the following steps:

[0007] S1. Multi-source data preprocessing and regional-feature-adaptive time window setting: preprocess multi-source meteorological data and dynamically set the time window τ based on meteorological element characteristics and regional features;

[0008] S2. Construction and training of the flow dependency model: Construct a deep learning flow dependency model with a CNN-LSTM hybrid architecture, and train it to learn the weight matrix W of the time dependency relationship between historical time and current time.

[0009] S3. Construction of Flow-Dependent 3D Variational Cost Function: Combining the weight matrix W and the concept of multi-scale assimilation, a flow-dependent 3D variational cost function is constructed.

[0010] S4. Solving variational equations and generating analysis field: The equations are solved using a multi-scale three-dimensional variational method, and the analysis field is obtained by superimposing multi-temporal information and updating it through a sliding time window.

[0011] Preferably, the adjustment range of the time window τ in step S1 is 3-10 days, with the τ value corresponding to high-variability meteorological elements and small areas being smaller, and the τ value corresponding to low-variability meteorological elements and large areas being larger.

[0012] Preferably, the flow dependency model in step S2 is a CNN-LSTM hybrid architecture, which has the ability to extract spatial features and capture long temporal dependencies, and outputs a temporal dependency weight matrix W.

[0013] Preferably, the flow-dependent three-dimensional variational cost function in step S3 introduces a weight matrix W, which dynamically reflects the contribution of different historical moments to the current assimilation process.

[0014] Preferably, in step S4, a sliding time window is used for cyclic assimilation, and the window length is consistent with the time window τ in step S1. The weight matrix W of the flow dependency model is updated in each round of assimilation.

[0015] Compared with related technologies, the three-dimensional variational data assimilation method based on a flow-dependent model provided by this invention has the following advantages:

[0016] This invention provides a three-dimensional variational data assimilation method based on a flow-dependent model. It introduces a deep learning flow-dependent model and accurately constructs a meteorological element association model in a small-scale time dimension through high-precision multi-feature extraction capabilities, replacing traditional static prior information. The new three-dimensional variational method has dynamic time adaptability.

[0017] By dynamically setting time windows for different regions and meteorological elements, rapid model training is achieved, which solves the adaptability problem of fixed time windows in traditional methods and further improves assimilation accuracy.

[0018] The construction flow relies on a three-dimensional variational cost function, but does not require complex adjoint modes. While taking into account the advantages of the four-dimensional variational time dimension, it significantly reduces computational costs and is easy to promote in business applications.

[0019] By using a sliding time window for cyclic assimilation and dynamic model updates, the time dependency is ensured to match the latest observational data, providing a more reliable initial field for numerical weather prediction and guaranteeing the accuracy and stability of the forecast results. Attached Figure Description

[0020] Figure 1 A flowchart illustrating the steps of a three-dimensional variational data assimilation method based on a flow-dependent model provided by this invention;

[0021] Figure 2 The flowchart illustrates the training and application of the flow dependency model provided by this invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] Please refer to the following: Figure 1 and Figure 2 ,in, Figure 1 A flowchart illustrating the steps of a three-dimensional variational data assimilation method based on a flow-dependent model provided by this invention; Figure 2 The flowchart illustrates the training and application of the flow dependency model provided by this invention.

[0024] A three-dimensional variational data assimilation method based on a flow-dependent model includes the following steps:

[0025] S1. Multi-source data preprocessing and regional-feature-adaptive time window setting: preprocess multi-source meteorological data and dynamically set the time window τ based on meteorological element characteristics and regional features;

[0026] S2. Construction and training of the flow dependency model: Construct a deep learning flow dependency model with a CNN-LSTM hybrid architecture, and train it to learn the weight matrix W of the time dependency relationship between historical time and current time.

[0027] S3. Construction of Flow-Dependent 3D Variational Cost Function: Combining the weight matrix W and the concept of multi-scale assimilation, a flow-dependent 3D variational cost function is constructed.

[0028] S4. Solving variational equations and generating analysis field: The equations are solved using a multi-scale three-dimensional variational method, and the analysis field is obtained by superimposing multi-temporal information and updating it through a sliding time window.

[0029] Multi-source meteorological observation data (satellite, ground station, radiosonde, radar, etc.) and numerical model background field data are collected, and spatiotemporal matching, outlier removal, and standardization are performed to ensure data consistency. Based on the characteristics of meteorological elements (temporal variability, spatial correlation) and regional features (spatial scale, underlying surface type), a time window τ (the time difference between historical and current moments) is dynamically set: for high-variability elements (precipitation, strong convective parameters): τ = 3-5 days, adapting to the rapid change characteristics of elements; for low-variability elements (sea level pressure, atmospheric humidity): τ = 7-10 days, preserving historical trend information; for small regions (urban, watershed scale): τ = 3-6 days, adapting to the heterogeneity of local meteorological fields; for large regions (intercontinental, global scale): τ = 6-10 days, ensuring the continuity of large-scale circulation.

[0030] A deep learning model (based on a CNN-LSTM hybrid architecture) with high-precision multi-feature extraction capability is constructed as a flow dependency model. The model input consists of meteorological element sequences (temperature, humidity, wind field, etc.), spatial location features, and underlying surface type features within a historical time period τ. The output is a weight matrix W of the temporal dependency relationship between historical moments and the current moment (reflecting the influence strength of different historical moments on the current state). The training dataset uses meteorological observation data and model background field data from the past 5-10 years. The samples are divided according to the time window τ corresponding to the region and the element. The loss function is to minimize the deviation between the observed field and the model prediction dependency relationship. The model parameters are optimized by gradient descent to ensure that the model can accurately capture the correlation patterns of meteorological elements in the small-scale time dimension. A robust training mechanism is introduced to improve the model's fault tolerance to outliers in the observed data by adding noisy samples and outlier perturbations.

[0031] Based on the concept of multi-scale assimilation, the three-dimensional variational objective function is decomposed into multiple time-state components. Combined with the time dependency weight matrix W output by the flow dependency model, a flow dependency three-dimensional variational cost function is constructed:

[0032]

[0033] Wherein, ε is the error signal vector (containing the error sequence ε_{t-τ}~ε_{t-1} within the current time ε_t and the historical time interval τ). The time-dependent background error covariance matrix, Let Y be the observation operator matrix, and Y be the observation information vector. The observation error covariance matrix is ​​used; the weight matrix W is dynamically updated through the flow dependency model, so that the cost function can reflect the contribution of different historical moments to the current assimilation process in real time, and help to correct the current motion state and direction of the elements.

[0034] A multi-scale three-dimensional variational method is employed to solve the flow-dependent variational equations. Through successive mesh coarsening, smoothing approximation, and correction, multi-scale information optimization is achieved, reducing computational complexity. The current background field and the observational information from the current and historical time periods τ are superimposed using a weight matrix W to obtain the final analysis field.

[0035]

[0036] in, For the current moment, analyze the field. The background scene at the current moment. The weights for the k-th historical moment in the output of the flow dependency model. This is the gain matrix at the corresponding time point. To obtain the new observation information at the corresponding time, a sliding time window (with the same window length as τ in step S1) is used for cyclic assimilation. During each round of assimilation, the flow dependency model is called again to update the weight matrix W, ensuring that the time dependency matches the latest observation data.

[0037] In step S1, the adjustment range of the time window τ is 3-10 days. The τ value corresponding to high-variability meteorological elements and small areas is relatively small, while the τ value corresponding to low-variability meteorological elements and large areas is relatively large.

[0038] The flow dependency model in step S2 is a CNN-LSTM hybrid architecture, which has the ability to extract spatial features and capture long temporal dependencies, and outputs a temporal dependency weight matrix W.

[0039] In step S3, the flow-dependent three-dimensional variational cost function introduces a weight matrix W, which dynamically reflects the contribution of different historical moments to the current assimilation process.

[0040] In step S4, a sliding time window is used for cyclic assimilation. The window length is the same as the time window τ in step S1. The weight matrix W of the flow dependency model is updated in each round of assimilation.

[0041] The working principle of the three-dimensional variational data assimilation method based on a flow-dependent model provided by this invention is as follows:

[0042] Data preprocessing stage:

[0043] Global satellite observation data (MODIS, AMSU, etc.), ground meteorological station observation data (temperature, humidity, precipitation), and WRF model background field data were collected from 2018 to 2028, with a spatial resolution of 0.25°×0.25° and a temporal resolution of 6 hours. Quality control algorithms were used to remove outliers exceeding reasonable thresholds. Spatiotemporal matching of different data sources was achieved through linear interpolation, and a unified format dataset was formed after standardization. For precipitation data (a high-variability element) in East my country (a small region), a time window of τ=4 days was set; for global sea level pressure data (a low-variability element), τ=8 days was set.

[0044] Training the stream dependency model:

[0045] A CNN-LSTM hybrid model was constructed: the CNN layer used three convolutional kernels (3×3, 5×5, and 3×3 respectively) to extract spatial features and small-scale temporal correlation features of meteorological elements; the LSTM layer used a two-layer bidirectional structure with 256 hidden units to capture long-term temporal dependencies; the output layer was a fully connected layer, with the output dimension corresponding to the time window τ (a 4×4 weight matrix was output when τ=4). The Adam optimizer was used during training, with an initial learning rate of 0.001, dynamically adjusted through a learning rate decay strategy. The training consisted of 100 epochs with a batch size of 32. After training, the model achieved an accuracy of over 92% in capturing small-scale temporal relationships.

[0046] Cost function construction and solution:

[0047] Based on the time window τ set in step S1 and the weight matrix W output by the flow-dependent model, a flow-dependent three-dimensional variational cost function is constructed. Taking precipitation assimilation in East China as an example, the weight coefficients for recent historical moments (t-1 to t-2 days) in the W matrix are 0.6 to 0.8, while the weight coefficients for distant historical moments (t-3 to t-4 days) are 0.2 to 0.4, reflecting the dominant role of recent observations in the current state. When using the multi-scale three-dimensional variational method, the grid coarsening ratio is 2:1, and the Gaussian filter radius is dynamically adjusted according to the grid scale (1 grid point for fine grid, 3 grid points for coarse grid) to ensure effective correction of small-scale signals while controlling the computational load.

[0048] Cyclic assimilation and accuracy verification:

[0049] The generated analysis field is used as the initial field for the WRF model, and cyclic assimilation is performed using a sliding time window (τ=4 days), with an assimilation update completed every 6 hours. Comparative experiments show that the root mean square error between the analysis field and the observation field of the method of this invention is reduced by more than 35% compared with the traditional static three-dimensional variational method, and the deviation in characterizing the location of small-scale heavy precipitation centers is reduced to within 5km; the computational cost is only 1 / 5 of that of the four-dimensional variational method, meeting the requirements of operational use.

[0050] Compared with related technologies, the three-dimensional variational data assimilation method based on a flow-dependent model provided by this invention has the following advantages:

[0051] By introducing a deep learning flow dependency model, a meteorological element association model on a small-scale time dimension is accurately constructed through high-precision multi-feature extraction capabilities, replacing the traditional static prior information. The new three-dimensional variational method has dynamic time adaptability.

[0052] By dynamically setting time windows for different regions and meteorological elements, rapid model training is achieved, which solves the adaptability problem of fixed time windows in traditional methods and further improves assimilation accuracy.

[0053] The construction flow relies on a three-dimensional variational cost function, but does not require complex adjoint modes. While taking into account the advantages of the four-dimensional variational time dimension, it significantly reduces computational costs and is easy to promote in business applications.

[0054] By using a sliding time window for cyclic assimilation and dynamic model updates, the time dependency is ensured to match the latest observational data, providing a more reliable initial field for numerical weather prediction and guaranteeing the accuracy and stability of the forecast results.

[0055] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A three-dimensional variational data assimilation method based on a flow-dependent model, characterized in that, The method comprises the following steps: S1, multi-source data preprocessing and regional-element adaptive time window setting: preprocessing multi-source meteorological data, dynamically setting a time window τ based on meteorological element characteristics and regional characteristics; S2, flow-dependent model construction and training: constructing a deep learning flow-dependent model with a CNN-LSTM hybrid architecture, and training a time-dependent relationship weight matrix W of historical time and current time; S3, construction of flow-dependent three-dimensional variational cost function: combining the weight matrix W and the multi-scale assimilation concept to construct a flow-dependent three-dimensional variational cost function; S4, solving of variational equation and generation of analysis field: solving the equation by using a multi-scale three-dimensional variational method, superimposing multi-time state information to obtain an analysis field, and updating by cyclic assimilation through a sliding time window.

2. The three-dimensional variational data assimilation method based on a flow-dependent model according to claim 1, wherein, The adjustment range of the time window τ in the step S1 is 3-10 days, and τ corresponding to high variability meteorological elements and small regions is small in value, and τ corresponding to low variability meteorological elements and large regions is large in value.

3. The three-dimensional variational data assimilation method based on a flow-dependent model according to claim 1, wherein, The flow-dependent model in the step S2 is a CNN-LSTM hybrid architecture, which has spatial feature extraction and long-time sequence dependence capturing capability, and the output is a time-dependent relationship weight matrix W.

4. The three-dimensional variational data assimilation method based on a flow-dependent model according to claim 1, wherein, The flow-dependent three-dimensional variational cost function in the step S3 introduces the weight matrix W, which dynamically reflects the contribution degree of different historical time to the current assimilation process.

5. The three-dimensional variational data assimilation method based on a flow-dependent model according to claim 1, wherein, In the step S4, a sliding time window is used for cyclic assimilation, the window length is consistent with the time window τ in the step S1, and the weight matrix W of the flow-dependent model is updated every round of assimilation.

Images