A cyclic assimilation method based on WRFDA coupled hot start
By introducing a cyclic assimilation and hot-start mechanism into the WRFDA assimilation system and using reinforcement learning algorithm to dynamically optimize parameters, the problem of relying on fixed parameters and lack of feedback mechanism in the WRFDA assimilation system is solved, and higher assimilation accuracy and forecast accuracy are achieved.
Patent Information
- Application Number
- CN202411749486.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-02
- Publication Date
- 2025-04-01
AI Technical Summary
The existing WRFDA assimilation system relies on fixed parameters and configuration in meteorological forecasting, resulting in errors in assimilation results, and the assimilation and forecasting process lacks an effective feedback mechanism, which limits the improvement of forecast accuracy.
The cyclic assimilation method based on WRFDA coupled hot start is adopted. By constructing the WRFDA assimilation system model and combining the reinforcement learning algorithm, the parameters in the assimilation process are dynamically optimized to achieve tight coupling and dynamic feedback between assimilation and forecasting.
By dynamically adjusting the parameters and configuration of the assimilation system, an optimized initial field is generated that is closer to the actual atmospheric state, which significantly improves the accuracy of assimilation and the accuracy and stability of forecasts.
Smart Images
Figure CN119738896A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of meteorological forecasting, and in particular to a cyclic assimilation method based on WRFDA coupled hot start. Background Art
[0002] In the field of weather forecasting, accurate and timely weather information is crucial to all aspects of social economy. In order to improve the accuracy of weather forecasts, data assimilation technology is widely used to integrate meteorological observation data from multiple sources into numerical weather forecast models, thereby generating an initial field that is closer to the actual atmospheric state. Among them, the WRFDA (Weather Research and Forecasting Data Assimilation) system, as an advanced data assimilation system, can effectively assimilate a variety of observation data, such as ground observations, satellite remote sensing data, radar data, etc., to provide high-quality initial conditions for numerical weather forecasts.
[0003] The current WRFDA assimilation system still faces some challenges in practical applications. First, the assimilation process often relies on fixed system parameters and configurations, which may not be optimal under complex and changeable meteorological conditions, resulting in certain errors in the assimilation results. Second, assimilation and forecasting are usually regarded as two independent processes, lacking an effective feedback mechanism to dynamically adjust the parameters in the assimilation process according to the forecast results, thus limiting the further improvement of forecast accuracy.
[0004] In addition, hot start technology is an effective means to improve the efficiency of numerical weather forecasting. By saving the state of the forecast model after it has run for a period of time and using it as the initial condition for the next forecast, it can significantly reduce the time required for the model to reach a stable state. However, traditional hot start technology is often separated from the assimilation process and does not fully utilize the optimized initial field information after assimilation, resulting in limited effects of hot start. Summary of the invention
[0005] The object of the present invention is to provide a cyclic assimilation method based on WRFDA coupled hot start to solve the problems raised in the above background technology.
[0006] To achieve the above object, the present invention provides the following technical solution: a cyclic assimilation method based on WRFDA coupled hot start, the method comprising:
[0007] Step 1: Construct the WRFDA assimilation system model, which is used to assimilate various meteorological observation data and generate an optimized initial meteorological field;
[0008] Step 2: Define the state space, covering the multi-dimensional information of the meteorological system, including temperature, humidity, wind speed, wind direction, air pressure and precipitation;
[0009] Step 3: Define the action space, including the adjustment actions for the WRFDA assimilation system parameters and forecast model configuration, as well as the triggering conditions of the hot start mechanism;
[0010] Step 4: Design a cyclic process of assimilation and forecasting, where each cycle includes assimilating new observation data using the WRFDA assimilation system, running the forecast model to generate forecast results, evaluating the forecast quality, and adjusting the parameters in the action space based on the evaluation results;
[0011] Step 5: Implement a hot start mechanism to save the mode state at the end of each cycle as the initial condition for the next cycle;
[0012] Step 6: Use reinforcement learning algorithm to train the WRFDA assimilation system model to optimize the parameter selection in the cyclic assimilation process.
[0013] Preferably, the state space further includes a time dimension for recording timestamps of observation data and forecast results.
[0014] Preferably, the training process of the deep reinforcement learning algorithm includes constructing a deep neural network as an approximator of the value function, and continuously updating the parameters of the neural network through experience replay and policy gradient descent methods to optimize the assimilation and prediction strategies.
[0015] Preferably, the steps of constructing the WRFDA assimilation system model by a reinforcement learning algorithm include:
[0016] Construct a deep neural network as an approximator of the value function, whose input is a representation of the state space and output is a value estimate of the corresponding action space;
[0017] Randomly initialize the parameters of the deep neural network, including weights and biases, as well as learning rates and discount factors;
[0018] During the training process, the states, actions, rewards, and next states generated by the interaction between the forecast system and meteorological observation data are stored in an experience replay buffer; when the buffer reaches a certain size, a batch of experiences are randomly extracted from it for network training.
[0019] Preferably, during the training process of the WRFDA assimilation system model, the gradient of the policy network is calculated by using samples extracted from the experience replay buffer using the gradient descent algorithm, and the parameters of the policy network are updated using the back propagation algorithm; after each iterative training, the mean square error MSE algorithm is used to calculate the loss of the model, and the loss function is calculated as follows:
[0020] L(θ)=(r+γmax a′ Q(s′,a′;θ′)-Q(s,a;θ))2
[0021] Where s is the current state, a is the action taken, r is the reward, s′ is the next state, γ is the discount factor, Q(s,a;θ) is the estimate of the state and action value function under the current strategy, and θ′ is the parameter of the target network.
[0022] Preferably, the output layer of the WRFDA assimilation system model is configured to generate assimilated meteorological element field data, which is represented in a gridded form, and each grid point contains a set of meteorological element values.
[0023] Preferably, the data format and structure of the output layer are consistent with the input observation data and background field data.
[0024] Preferably, during the training process of the deep reinforcement learning algorithm, a balance mechanism of exploration and exploitation is further introduced, and the balance mechanism adopts an ε-greedy strategy.
[0025] Preferably, the method further comprises: performing quality control on the inputted various meteorological observation data, including encoding processing and standardization processing on the inputted data.
[0026] Preferably, the encoding processing method for the input data includes: performing one-hot encoding conversion on discrete features;
[0027] The method of standardizing the input data includes: using the Z-score standardization method to standardize the continuous features, specifically: subtracting the mean of each numerical data and dividing it by its standard deviation, so that the processed data conforms to the standard normal distribution. The formula for standardization is:
[0028] =(-) / σ, where is the original data, is the mean of the original data, is the standard deviation of the original data, and is the standardized data.
[0029] Compared with the prior art, the present invention has the following beneficial effects:
[0030] By constructing the WRFDA assimilation system model and combining the reinforcement learning algorithm to dynamically optimize the parameters in the assimilation process, the present invention can automatically adjust the assimilation system parameters and configurations according to different meteorological conditions and observation data characteristics, thereby generating an optimized initial field that is closer to the actual atmospheric state. This effectively solves the assimilation error problem caused by the traditional WRFDA assimilation system's reliance on fixed parameters and configurations, and significantly improves the accuracy of assimilation. The present invention closely combines the assimilation and forecasting processes and designs a cyclic assimilation and forecasting process. In this process, each cycle will dynamically adjust the parameters of the assimilation system based on the evaluation results of the forecast quality, realizing effective feedback between assimilation and forecasting. This mechanism not only improves the accuracy of the forecast, but also enhances the stability of the forecast, making the forecast results more reliable.
[0031] By implementing a hot start mechanism, the present invention saves the mode state at the end of each cycle as the initial condition for the next cycle. This greatly reduces the time required for the forecast model to reach a stable state and improves the efficiency of the forecast. At the same time, because the hot start mechanism is closely coupled with the assimilation process, the optimized initial field information after assimilation is fully utilized, further improving the accuracy and efficiency of the forecast. The present invention defines a state space containing multi-dimensional meteorological information and an adjustable action space, so that the system can flexibly adjust the assimilation system parameters and forecast model configuration according to actual needs. This design enhances the adaptability and flexibility of the system, enabling it to cope with various complex and changeable meteorological conditions and provide more comprehensive and accurate support for meteorological forecasting. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 This is a working principle diagram of a cyclic assimilation method based on WRFDA coupled hot start according to the present invention;
[0033] Figure 2 This is the training flow chart of the WRFDA assimilation system model. DETAILED DESCRIPTION
[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0035] See also Figure 1-2 The present invention provides a technical solution: a cyclic assimilation method based on WRFDA coupled hot start, the method comprising:
[0036] Step 1: Construct the WRFDA assimilation system model
[0037] Construct a WRFDA assimilation system model that has the ability to assimilate a variety of meteorological observation data, including but not limited to ground observation data, satellite remote sensing data, and radar data. Through the algorithm of the WRFDA assimilation system, these observation data are integrated into the numerical weather forecast model to generate an optimized initial meteorological field. This initial meteorological field will serve as the starting point for subsequent forecasts to ensure the accuracy of the forecast.
[0038] Step 2: Define the state space
[0039] Define the state space, which covers the multi-dimensional information of the meteorological system. Specifically, the state space includes key meteorological elements such as temperature, humidity, wind speed, wind direction, air pressure and precipitation. The state space also includes the time dimension, which is used to record the timestamps of observation data and forecast results.
[0040] Step 3: Define the action space
[0041] After defining the state space, the action space is further defined. The action space contains the adjustment actions for the WRFDA assimilation system parameters and forecast model configuration, such as adjusting the assimilation time window, changing the background error covariance matrix, etc. In addition, the action space also includes the triggering conditions of the hot start mechanism, such as setting a specific time interval or forecast error threshold to determine when to start the hot start process.
[0042] Step 4: Design the assimilation and forecasting cycle
[0043] A cyclic process of assimilation and forecasting is designed, which is the core part of the present invention. Each cycle includes the following steps: first, the new observation data is assimilated using the WRFDA assimilation system to update the initial meteorological field; then, the forecast model is run to generate forecast results; then, the forecast quality is evaluated, such as calculating the forecast error, comparing the observation data, etc.; finally, the parameters in the action space are adjusted according to the evaluation results to optimize the assimilation and forecasting process. This cyclic process ensures the close coupling and dynamic feedback between assimilation and forecasting.
[0044] Step 5: Implement a hot start mechanism
[0045] At the end of each cycle, a hot start mechanism is implemented. Specifically, the state of the current forecast model, including the distribution and changes of each meteorological element, is saved. These saved states will be used as the initial conditions for the next cycle, thereby significantly reducing the time required for the model to reach a stable state and improving forecast efficiency.
[0046] Step 6: Use reinforcement learning algorithm to train the WRFDA assimilation system model
[0047] Finally, the reinforcement learning algorithm is used to train the WRFDA assimilation system model. The reinforcement learning algorithm simulates the cycle of assimilation and forecasting, and continuously tries different parameter selections and action combinations to maximize the accuracy of the forecast or minimize the forecast error. During the training process, the algorithm adjusts the parameter selection strategy according to the feedback results of each cycle (i.e., forecast quality assessment), thereby gradually optimizing the assimilation process. Through long-term training and learning, the WRFDA assimilation system model can automatically adapt to complex and changing meteorological conditions and improve the accuracy and stability of forecasts.
[0048] The present invention will be further described below in conjunction with Examples 1 to 3:
[0049] Embodiment 1:
[0050] In this invention, a deep reinforcement learning algorithm is used to train the WRFDA assimilation system model to optimize the parameter selection in the assimilation and forecasting process. The specific implementation steps are as follows:
[0051] Build a deep neural network: Build a deep neural network as an approximator of the value function. The input of the network is a representation of the state space, which not only contains multi-dimensional information of the meteorological system (such as temperature, humidity, wind speed, wind direction, air pressure and precipitation), but also includes a time dimension for recording the timestamps of observation data and forecast results. The output of the network is the value estimate of the corresponding action space, that is, the cumulative reward expected after taking a certain action.
[0052] Initialize network parameters: Randomly initialize the parameters of the deep neural network, including weights and biases, as well as learning rates and discount factors. These parameters will affect the training process and final performance of the neural network. The learning rate determines the step size of the parameter update, while the discount factor is used to calculate the present value of future rewards.
[0053] Experience replay mechanism: During the training process, the states, actions, rewards, and next states generated by the interaction between the forecast system and the meteorological observation data are stored in an experience replay buffer. This buffer is used to store recent experience samples so that a batch of samples can be randomly selected for learning during training, thereby breaking the time correlation between samples and improving the stability and efficiency of training. When the buffer reaches a certain size, a batch of experience is randomly selected from it for network training.
[0054] Policy network training:
[0055] The training is performed using samples drawn from the experience replay buffer using the gradient descent algorithm. For each sample, the gradient of the policy network is calculated, which reflects the rate of change of the expected reward after taking an action under the current policy. Then, the parameters of the policy network are updated through the back-propagation algorithm to maximize the expected reward.
[0056] After each iteration of training, the mean square error (MSE) algorithm is used to calculate the model loss. The loss function is used to measure the difference between the estimated state and action value function and the target value under the current strategy. The calculation formula of the loss function is:
[0057] L(θ)=(r+γmax a′ Q(s′,a′;θ′)-Q(s,a;θ)) 2
[0058] Among them, s is the current state, a is the action taken, r is the reward, s′ is the next state, γ is the discount factor, Q(s,a;θ) is the estimate of the state and action value function under the current strategy, and θ′ is the parameter of the target network (the target network is used to stabilize the training process, and its parameters are regularly copied from the main network).
[0059] Through continuous iterative training, the parameters of the deep neural network will gradually converge, enabling the WRFDA assimilation system model to automatically adapt to complex and changeable meteorological conditions and optimize the assimilation and forecasting strategies.
[0060] Embodiment 2:
[0061] In the construction of the WRFDA assimilation system model, the output layer is specially configured to generate assimilated meteorological element field data. These data are represented in a gridded form, covering the entire forecast area. Each grid point contains a complete set of meteorological element values, such as temperature, humidity, wind speed, wind direction, air pressure, and possible precipitation. This gridded representation facilitates the subsequent meteorological forecast model to perform spatial analysis and prediction, and is consistent with the format and structure of the input observation data and background field data, ensuring data consistency and processability.
[0062] In the training process of deep reinforcement learning algorithms, in order to balance the relationship between exploring new strategies and utilizing known strategies, the ε-greedy strategy is introduced as a balancing mechanism between exploration and utilization. The specific implementation is as follows:
[0063] Exploration phase: In the early stages of training or when the strategy performs poorly, the algorithm randomly selects an action with a certain probability ε (called the exploration rate) instead of always selecting the action with the highest estimated value. This helps the algorithm explore unknown action spaces and may discover better strategies.
[0064] Utilization phase: As training progresses and the strategy is continuously optimized, the algorithm selects the action with the highest estimated value at the moment with a probability of 1-ε, that is, it uses the known optimal or approximately optimal strategy to obtain the maximum cumulative reward.
[0065] Dynamic adjustment of ε: In order to balance exploration and exploitation, the value of ε is usually gradually reduced as training progresses. In the early stages of training, a larger ε value encourages the algorithm to explore more; in the later stages of training, a smaller ε value encourages the algorithm to make more use of known strategies to stabilize and improve prediction performance.
[0066] Taking temperature forecast as an example, suppose the WRFDA assimilation system model is assimilating and forecasting temperature in a certain area. During the training process, the algorithm first assimilates a variety of observation data (such as ground station observations, satellite remote sensing data, etc.) through the WRFDA system to generate the assimilated initial temperature field. Then, based on the current state (including the assimilated temperature field and other meteorological elements), the algorithm uses a deep neural network to estimate the value of different actions (such as adjusting assimilation parameters, changing model configuration, etc.).
[0067] In the exploration phase, the algorithm may randomly select an action with a certain probability ε, such as trying a previously unused assimilation parameter combination. After executing the action, the algorithm evaluates the quality of the forecast results (such as calculating the error by comparing with the actual observation data) and stores this experience (state, action, reward, next state) in the experience replay buffer.
[0068] In the utilization phase, the algorithm selects the action with the highest estimated value based on the current state, that is, the action that is considered to bring the largest cumulative reward. As training progresses, the algorithm continuously updates the parameters of the neural network through experience replay and gradient descent methods to optimize the assimilation and prediction strategies.
[0069] At the same time, by dynamically adjusting the value of ε, the algorithm finds a balance between exploration and utilization. In the early stage of training, a larger ε value enables the algorithm to explore different action spaces extensively; while in the later stage of training, a smaller ε value ensures that the algorithm can stably use the known optimal strategy for temperature prediction.
[0070] Embodiment 3:
[0071] Before inputting meteorological observation data into the WRFDA assimilation system, they must first undergo strict quality control, including two main steps: encoding processing and standardization processing.
[0072] For discrete features in the input data (such as observation station type, observation instrument type, etc.), one-hot encoding is used for conversion. One-hot encoding is a method that converts each value of a discrete feature into a one-hot vector. Each bit in the vector represents a possible value, and only when the value corresponding to the bit appears, the bit is 1, and the rest of the bits are 0. In this way, discrete features can be converted into numerical features that can be processed by machine learning models.
[0073] For example, suppose there are three types of observation stations: ground station, high-altitude station, and ocean station, represented by 1, 2, and 3 respectively. For a certain observation data, if its observation station type is a ground station, then after one-hot encoding, the feature will be converted to [1,0,0]; if it is an high-altitude station, it will be converted to [0,1,0]; if it is an ocean station, it will be converted to [0,0,1].
[0074] For continuous features in the input data (such as temperature, humidity, wind speed, etc.), the Z-score standardization method is used for processing. Z-score standardization is a method of scaling data so that it conforms to the standard normal distribution (mean is 0 and standard deviation is 1). Specifically, for each continuous feature, its mean and standard deviation are calculated, and then each numerical data is subtracted from the mean and divided by its standard deviation to obtain the standardized data.
[0075] The formula for standardization is: =(-) / σ, where is the original data, is the mean of the original data, is the standard deviation of the original data, and is the standardized data.
[0076] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0077] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A cyclic assimilation method based on WRFDA coupled hot start, characterized in that: The method comprises: Step 1: Construct the WRFDA assimilation system model, which is used to assimilate various meteorological observation data and generate an optimized initial meteorological field; Step 2: Define the state space, covering the multi-dimensional information of the meteorological system, including temperature, humidity, wind speed, wind direction, air pressure and precipitation; Step 3: Define the action space, including the adjustment actions for the WRFDA assimilation system parameters and forecast model configuration, as well as the triggering conditions of the hot start mechanism; Step 4: Design a cyclic process of assimilation and forecasting, where each cycle includes assimilating new observation data using the WRFDA assimilation system, running the forecast model to generate forecast results, evaluating the forecast quality, and adjusting the parameters in the action space based on the evaluation results; Step 5: Implement a hot start mechanism to save the mode state at the end of each cycle as the initial condition for the next cycle; Step 6: Use reinforcement learning algorithm to train the WRFDA assimilation system model to optimize the parameter selection in the cyclic assimilation process.
2. A cyclic assimilation method based on WRFDA coupled hot start according to claim 1, characterized in that: The state space further includes a time dimension for recording timestamps of observation data and forecast results.
3. A cyclic assimilation method based on WRFDA coupled hot start according to claim 1, characterized in that: The training process of the deep reinforcement learning algorithm includes constructing a deep neural network as an approximator of the value function, and continuously updating the parameters of the neural network through experience replay and policy gradient descent methods to optimize the assimilation and prediction strategies.
4. A cyclic assimilation method based on WRFDA coupled hot start according to claim 3, characterized in that: The steps to construct the WRFDA assimilation system model through reinforcement learning algorithm include: Construct a deep neural network as an approximator of the value function, whose input is a representation of the state space and output is a value estimate of the corresponding action space; Randomly initialize the parameters of the deep neural network, including weights and biases, as well as learning rates and discount factors; During the training process, the states, actions, rewards, and next states generated by the interaction between the forecast system and meteorological observation data are stored in an experience replay buffer; when the buffer reaches a certain size, a batch of experiences are randomly extracted from it for network training.
5. A cyclic assimilation method based on WRFDA coupled hot start according to claim 4, characterized in that: In the training process of the WRFDA assimilation system model, the gradient descent algorithm is used to extract samples from the experience replay buffer to calculate the gradient of the policy network, and the back propagation algorithm is used to update the parameters of the policy network; after each iterative training, the mean square error MSE algorithm is used to calculate the loss of the model, and the calculation formula of the loss function is: L(θ)=(r+γmax a′ Q(s′,a′;θ′)-Q(s,a;θ)) 2 Where s is the current state, a is the action taken, r is the reward, s′ is the next state, γ is the discount factor, Q(s,a;θ) is the estimate of the state and action value function under the current strategy, and θ′ is the parameter of the target network.
6. A cyclic assimilation method based on WRFDA coupled hot start according to claim 5, characterized in that: The output layer of the WRFDA assimilation system model is configured to generate assimilated meteorological element field data, which is represented in a gridded form, and each grid point contains a set of meteorological element values.
7. A cyclic assimilation method based on WRFDA coupled hot start according to claim 6, characterized in that: The data format and structure of the output layer are consistent with the input observation data and background field data.
8. A cyclic assimilation method based on WRFDA coupled hot start according to claim 4, characterized in that: During the training process of the deep reinforcement learning algorithm, a balance mechanism of exploration and exploitation is further introduced, and the balance mechanism adopts an ε-greedy strategy.
9. A cyclic assimilation method based on WRFDA coupled hot start according to claim 1, characterized in that: The method also includes: performing quality control on the inputted multiple meteorological observation data, including encoding processing and standardization processing on the inputted data.
10. A cyclic assimilation method based on WRFDA coupled hot start according to claim 9, characterized in that: Methods for encoding input data include: one-hot encoding conversion of discrete features; The method of standardizing the input data includes: using the Z-score standardization method to standardize the continuous features, specifically: subtracting the mean of each numerical data and dividing it by its standard deviation, so that the processed data conforms to the standard normal distribution. The formula for standardization is: =(-) / σ, where is the original data, is the mean of the original data, is the standard deviation of the original data, and is the standardized data.