A precipitation response deduction method based on physical constraint and job disturbance guidance

Abstract

This invention relates to a precipitation response extrapolation method based on physical constraints and operational disturbance guidance, comprising: acquiring big data of meteorological information, ground-fused precipitation information, and artificial rain enhancement operation information according to a preset sampling strategy; performing big data processing to generate an input tensor based on the meteorological information and operational information; generating a first label based on the actual precipitation amount and a second label based on the actual precipitation amount and the precipitation amount without rain enhancement operations; training a model based on the input tensor and the first label, and calculating a first loss function; calculating a second loss function based on the input tensor and a preset physical constraint model; determining the first and second precipitation prediction amounts based on the input tensor and the second label; determining a third loss function; optimizing the model based on the first, second, and third loss functions until the model converges; and outputting the precipitation increment of the rain enhancement operation when the model converges.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a precipitation response extrapolation method based on physical constraints and operational disturbance guidance. Background Technology

[0002] In meteorology, artificial rain enhancement operations, such as the deployment of rain-enhancing materials via aircraft, rockets, or ground-based equipment, are a common weather intervention aimed at promoting precipitation by altering local atmospheric conditions. However, evaluating the effectiveness of artificial rain enhancement operations remains a complex challenge. Traditional evaluation methods typically rely on historical data, meteorological simulations, and empirical rules, lacking an intelligent tool capable of predicting and assessing operational effectiveness in real time and quantitatively.

[0003] The effectiveness of artificial rain enhancement operations is influenced by a variety of factors, including meteorological conditions such as temperature, humidity, and wind speed; operational parameters such as operation time, location, operation method, and dosage; and regional and seasonal factors. While existing numerical weather prediction models, such as Cyclic Prefix Based Modulation-Assisted Extended Space Orthogonal (CMA-MESO) and Weather Research and Forecasting (WRF), and precipitation observation data such as the China Multi-Model Ensemble Prediction System (CMPAS), provide a wealth of meteorological and precipitation data, these data typically cannot directly provide a quantitative assessment of the operation's impact. Furthermore, traditional models have limitations in handling complex nonlinear weather disturbances and causal relationships, particularly in accurately identifying and inferring precipitation changes caused by operational disturbances.

[0004] Therefore, establishing an intelligent evaluation system that integrates meteorological data, operational disturbance information, and causal inference models has become the key to improving the efficiency and scientific rigor of artificial rain enhancement operations. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a precipitation response extrapolation method based on physical constraints and operational disturbance guidance, thereby solving the problems existing in the prior art.

[0006] To achieve the above objectives, this invention provides a precipitation response extrapolation method based on physical constraints and operational disturbance guidance, comprising:

[0007] According to a preset sampling strategy, meteorological information, surface-fused precipitation information, and artificial rain enhancement operation information are acquired; the meteorological information includes precipitation before rain enhancement operations; the surface-fused precipitation information includes the actual precipitation during rain enhancement operations.

[0008] Based on the meteorological information and the operational information, an input tensor is generated; the input tensor includes a time dimension, a spatial dimension, the number of meteorological element channels, and a disturbance channel;

[0009] A first tag is generated based on the actual precipitation, and a second tag is generated based on the actual precipitation and the precipitation without rain enhancement operations.

[0010] Based on the input tensor and the first label, the model is trained, and the first loss function is calculated.

[0011] The second loss function is calculated based on the input tensor and the preset physical constraint model;

[0012] Based on the input tensor and the second label, the first precipitation forecast during the operation and the second precipitation forecast without rain enhancement operation are determined respectively.

[0013] Calculate the first loss term for the first precipitation forecast, and calculate the second loss term for the second precipitation forecast;

[0014] Based on the first precipitation forecast and the second precipitation forecast, determine the first difference;

[0015] Based on the first difference and the second label, determine the causal supervision loss term;

[0016] Based on the first loss term, the second loss term, and the causal supervision loss term, a third loss function is determined;

[0017] The model is optimized based on the first loss function, the second loss function, and the third loss function until the model converges.

[0018] When the model converges, the output shows the increase in precipitation from the rain enhancement operation.

[0019] By applying the precipitation response extrapolation method based on physical constraints and operational perturbation provided in this invention, a spatiotemporal feature extraction network based on ConvLSTM is used to capture spatial correlations through convolutional operations and to model temporal dependencies using the LSTM mechanism, achieving a deep understanding of atmospheric evolution. The physical constraint module introduces fundamental meteorological principles such as mass conservation, energy balance, and momentum constraints as hard constraints to ensure that the model output conforms to physical laws. The causal extrapolation framework accurately separates the effects of operational perturbation and natural precipitation through parallel training of intervention and control branches, ultimately outputting a precipitation increment ΔP with causal explanation. In terms of model optimization, a multi-task joint loss function is adopted, comprehensively considering prediction accuracy, regression loss, causal consistency, causal response loss, physical rationality, and constraint loss. Furthermore, the AdamW optimizer and gradient pruning technique are innovatively combined to effectively improve the stability and convergence speed of model training.

[0020] Compared to traditional methods, this invention offers three significant advantages: First, through spatiotemporal dynamic modeling and precise perturbation coding, it achieves a refined assessment of the impact of artificial rain enhancement operations; second, the introduction of physical constraint mechanisms ensures the meteorological rationality of the forecast results; and finally, the causal inference framework provides interpretable quantitative indicators of operational effectiveness. This method offers a real-time, accurate, and interpretable operational effectiveness evaluation mechanism, overcoming the limitations of traditional methods that rely on simulation and speculation. It can provide reliable decision support for meteorological departments and optimize rain enhancement operation strategies. By comprehensively utilizing meteorological forecast data, operational information, causal inference, and physical constraints, this invention has broad application prospects in the field of artificial rain enhancement operation evaluation, particularly in disaster response, environmental protection, and meteorological intervention. Attached Figure Description

[0021] Figure 1 This is one of the flowcharts of a precipitation response extrapolation method based on physical constraints and operational disturbances provided in an embodiment of the present invention.

[0022] Figure 2 The second schematic diagram of the precipitation response extrapolation method based on physical constraints and operational disturbances provided in this embodiment of the invention;

[0023] Figure 3 This is a schematic diagram of the training sequence of key modules of the precipitation response extrapolation model provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0025] Figure 1 This is a schematic diagram of the precipitation response extrapolation method based on physical constraints and operational disturbance guidance provided in an embodiment of the present invention. The following is in conjunction with... Figure 1 The technical solution of the present invention will be described with reference to specific embodiments. Figure 1 As shown, this application includes the following steps:

[0026] Step 101: According to the preset sampling strategy, acquire meteorological information, surface-fused precipitation information, and artificial rain enhancement operation information; the meteorological information includes precipitation before rain enhancement operations; the surface-fused precipitation information includes the actual precipitation during rain enhancement operations.

[0027] This application collected various big data, including ground-based fusion precipitation information and artificial rain enhancement operation information, and processed the big data. The meteorological information, specifically CMA-MESO, is a mesoscale numerical weather prediction model widely used in weather forecasting, climate analysis, and disaster early warning. It is mainly used to simulate and forecast small-scale weather phenomena in the atmosphere, such as thunderstorms, turbulence, and local storms, and has high spatial and temporal resolution. CMA-MESO's spatial resolution can reach 3 kilometers, giving it an advantage in capturing local weather events such as heavy rain and thunderstorms.

[0028] CMA-MESO data incorporates various meteorological variables, such as temperature, humidity, air pressure, wind speed, and precipitation, and is generated by a high-resolution numerical weather prediction (NWP) model. Specifically, the CMA-MESO model combines multi-source information, including surface meteorological observations, satellite observations, and radar echoes, using data assimilation techniques to provide accurate weather forecasts. Through the fusion of these multi-source data, CMA-MESO can provide accurate short-term weather forecasts, and is particularly suitable for simulating local weather events.

[0029] In the impact assessment of artificial rain enhancement operations in this application, CMA-MESO model forecast data serves as a reference for natural precipitation under no-operation conditions. This data helps the model learn how meteorological conditions lead to the formation and evolution of precipitation under conditions without any artificial rain enhancement intervention. Specifically, variables such as precipitation, humidity, wind speed, and temperature in CMA-MESO reflect the precipitation background under natural meteorological conditions before the operation, providing a benchmark for assessing the impact of artificial rain enhancement operations.

[0030] When using CMA-MESO data, the model input includes meteorological background data for a certain period before and after the operation, with particular focus on precipitation and other precipitation-related meteorological variables such as temperature, humidity, wind speed (including vertical wind speed and air pressure). This data represents the background conditions before artificial rain enhancement. CMA-MESO data is presented in a spatiotemporal grid format, with meteorological data at each time step serving as model input to simulate the atmospheric environment before the operation. These meteorological elements are processed through convolutional layers and combined with perturbation data to provide the meteorological background before the operation.

[0031] CMPAS integrates various observation and forecast information from ground-based automatic weather stations, satellite remote sensing, weather radar, and weather forecasting models, providing high spatiotemporal resolution precipitation observation data nationwide. CMPAS's advantages lie in its high spatial resolution (e.g., 1 km) and high temporal resolution (e.g., 1 hour), enabling it to accurately capture precipitation changes under different meteorological conditions.

[0032] CMPAS data overcomes the limitations of single observation methods, such as relying solely on satellites or radar, by fusing multi-source precipitation observation data. For example, ground-based weather stations provide reliable measured precipitation data, but their spatial coverage is limited; while satellite remote sensing and radar can provide extensive spatial coverage and continuous dynamic observations, but they are subject to certain errors. By fusing these data, CMPAS can provide more comprehensive and accurate precipitation data, suitable for large-scale precipitation assessment and analysis.

[0033] In this application, CMPAS data is used as the actual precipitation data following artificial rain enhancement operations. Specifically, the CMPAS dataset provides observational data of actual precipitation after the operations, which is used as a supervisory signal to help the model learn the effectiveness of the rain enhancement operations during training. CMPAS not only provides precipitation data nationwide, but also ensures high accuracy of the precipitation data through data assimilation techniques, especially in areas with high precipitation intensity.

[0034] The model training process utilizes CMPAS data to supervise the evaluation of operational effectiveness. By comparing with CMA-MESO data, the model can identify the precipitation increment ΔP before and after the operation, and adjust the predicted precipitation effect based on CMPAS data. The combination of CMPAS data with artificial rain enhancement operation information, such as operation methods and dosage, provides the model with concrete, real-world observational results regarding the specific impacts of rain enhancement and operational interventions.

[0035] Artificial rain enhancement operation disturbance information refers to detailed information used to record and analyze artificial rain enhancement operations conducted within China. This data is crucial for evaluating the effectiveness of artificial rain enhancement operations, optimizing operational strategies, and providing a scientific basis for meteorological decision-making. Specifically, artificial rain enhancement operation information data includes basic operation information and operation process information. The basic operation information describes the basic background of the operation, including the time, location, and method of operation. The operation process information details the process of the rain enhancement operation, the types of equipment involved in the operation (e.g., aircraft, rockets, ground devices), the rain enhancement technologies used (e.g., cloud seeding, artificial ice cores), and the weather conditions during the operation (e.g., temperature, humidity, wind speed).

[0036] The operational method and dosage affect the intensity of the operation. This operational information is input into the ConvLSTM through a specific set of perturbation channels as part of the model, helping the model learn the impact of rain enhancement operations on precipitation.

[0037] The preset sampling strategies include event-driven sampling and control sample sampling. Event-driven sampling uses the moment of the artificial rain enhancement operation as the center of the sampling window and extracts data before and after the operation according to a preset step size. Control sample sampling collects data when there is no artificial rain enhancement operation. Since artificial rain enhancement events are discrete in time, the sampling period needs to be determined based on the timing of the operation and meteorological conditions when constructing the training set. To improve the training effect of the model, the following sampling strategy is adopted:

[0038] (1) Event-driven sampling

[0039] During sampling, the focus is on collecting data around the time of the artificial rain enhancement operation. This sampling method ensures that the training set contains sufficient positive samples, i.e., samples that include rain enhancement operation interventions. By selecting the operation event time as the center of the sampling window, the time periods before and after the operation are extracted as sample data. For example, data from 1 hour before the operation, the operation time, and 4 hours after the operation are selected as the training window. Within this window, CMA-MESO meteorological data are used as input, and CMPAS precipitation data are used as labels.

[0040] (2) Sampling of control samples

[0041] In addition, to enable the model to learn the evolution of natural precipitation under no-operation conditions, a certain proportion of control samples, i.e., precipitation processes without operations, are also included. These control samples can be selected from time periods in the same season and region as the operated samples, with similar weather backgrounds, but confirming that no rain enhancement operations were carried out. By matching environmental conditions, the model can learn to distinguish the differences in precipitation development between the "operation-enabled" and "operation-disabled" scenarios. The length T of the sampling period can be determined based on the lifetime of the precipitation system and the duration of the operation's impact. For example, if the typical convective cell lifetime is 1-3 hours, then a T of 3 hours can basically cover the entire process; stratiform precipitation may last longer, so a longer sequence or multiple sample segments can be used.

[0042] Step 102: Generate an input tensor based on the meteorological data and the operational information; the input tensor includes a time dimension, a spatial dimension, the number of meteorological element channels, and a disturbance channel;

[0043] Specifically, the meteorological information includes multiple meteorological variables, through... The meteorological variables are standardized; where X is the meteorological variable, μ is the mean of the meteorological variable, and σ is the standard deviation of the meteorological variable.

[0044] The standardized meteorological information is processed into a four-dimensional tensor [T,H,W,C]. meteo ], where T represents the time step, H represents the height of the spatial grid, W represents the width of the spatial grid, and C meteo The number of channels representing meteorological variables;

[0045] Align the time of the operation information with the timestamps of the meteorological variables;

[0046] Map the time range from the start time to the end time of the task information to the effective time step in the perturbation tensor;

[0047] Convert the dosage or intensity of the application in the operation information into perturbation values;

[0048] according to Construct a perturbation field and represent the operation perturbation channel through a spatial mask or perturbation vector to generate an operation perturbation tensor [T,H,W,1]; where Dt0 represents the perturbation field, S represents the perturbation intensity value, and S∈[0,1];

[0049] Based on the four-dimensional tensor and the operation perturbation tensor, generate the input tensor [T,H,W,C]. meteo +1], where C meteo It represents the channel number of the meteorological variable, with +1 indicating the disturbance channel.

[0050] The process will now be described in detail:

[0051] For all meteorological data input into the model (such as temperature, humidity, precipitation, etc.), Z-score standardization is first required, and the formula is as follows:

[0052]

[0053] Here, X represents the original meteorological data, and μ and σ are the mean and standard deviation of the variable, respectively. Through standardization, the data of all meteorological variables are transformed into a standard normal distribution with zero mean and unit variance, thereby avoiding unnecessary influences caused by inconsistencies in the units of measurement between different meteorological variables.

[0054] For each sample, the input data consists of the following parts:

[0055] Meteorological data (CMA-MESO): Includes meteorological elements such as temperature, humidity, wind speed, and air pressure. The spatial resolution of the input data is 3 km, and the temporal resolution is 1 hour. The meteorological data is formatted as [T, H, W, C]. meteo A four-dimensional tensor, where T represents the time step, H and W are the height and width of the spatial grid, respectively, and C... meteo It refers to the number of channels for meteorological variables, such as temperature, humidity, and wind speed.

[0056] Operation Information: Operation information includes operation time, location, operation method, dosage, etc. This information is converted into perturbation channels and inserted into the input tensor at the operation time. The perturbation channels are represented by spatial masks or perturbation vectors, and the specific spatial location and intensity are encoded according to the actual situation of the operation. This part of the data has the shape [T,H,W,1], that is, one perturbation channel per time step, and grid points near the operation location are assigned non-zero values ​​at the operation time. Each artificial rain enhancement operation includes:

[0057] Assignment time (e.g., 2021-07-20 12:00)

[0058] Work location (e.g., latitude and longitude range [lon1,lon2], [lat1,lat2])

[0059] Operational methods (e.g., aerial seeding, rocket bombing)

[0060] Dosage or intensity (e.g., silver iodide: 100g)

[0061] The solution is as follows:

[0062] a) First, align the operation time with the time step t of the meteorological data, assuming that the data has 1 time step per hour;

[0063] b) Map the time range from the start to the end of the task to the effective time step in the perturbation tensor;

[0064] c) Convert information such as dosage or intensity into perturbation values, such as intensity value S∈[0,1], to facilitate encoding into the perturbation channel.

[0065] Spatial perturbation mask generation

[0066] a) At operation time step t0, the corresponding perturbation channel D[t0,:,:,0] needs to be constructed using a spatial mask:

[0067] b) First, map the latitude and longitude range of the operation to the spatial grid points (H,W) of the meteorological field, and complete the task using grid interpolation.

[0068] c) For the area affected by the disturbance (i.e., the grid point corresponding to the geographical location of the operation), fill in a non-zero value S; for other areas, fill in 0.

[0069] The perturbation field can be constructed using a masking method:

[0070]

[0071] After constructing the perturbation tensor, it is then stitched together with meteorological data:

[0072] X input [T,H,W,C] meteo +1]=concat(X meteo [T,H,W,C meteo ],D[T,H,W,0])

[0073] The shape of the tensor after splicing is: X input [T,H,W,C meteo +1]. Where [T,H,W] are the time and space dimensions, and C is the C... meteo It represents the number of meteorological element channels, and +1 represents the disturbance channel.

[0074] Assume the CMA-MESO meteorological data dimensions are: T = 8, representing 8 hours; H = 100, representing height of 100; W = 100, representing width of 100; and C_meteo = 8, representing 8 meteorological variable channels.

[0075] The operation occurs at t=5, at location (50,50), and affects a 7×7 area. The operation intensity is S=0.8.

[0076] The final concatenated input tensor has a shape of [8, 100, 100, 9], with 9 channels including meteorological variables and operational disturbance information, and the operational mode channel serving as an additional channel. This input data, after being processed by a ConvLSTM model, can predict precipitation changes before and after the operation, thereby evaluating the effectiveness of artificial rain enhancement operations.

[0077] To ensure the spatiotemporal consistency of the training data, all meteorological data and precipitation labels need to be temporally and spatially aligned. First, ensure that the time step of CMA-MESO is synchronized with the time step of the precipitation data from CMPAS. If there are temporal discrepancies, temporal alignment should be performed through interpolation or nearest-neighbor sampling. For spatial alignment, ensure that the meteorological data at each grid point is consistent with the precipitation data at its corresponding location. For the timing of rain enhancement operations, perturbation information is inserted into the input tensor, and this perturbation information is ensured to be aligned with the precipitation time step, thereby ensuring that the model can "sense" the intervention at the correct time step.

[0078] The construction of the training set needs to ensure sufficient representativeness, covering different weather conditions, different types of rain enhancement operations, and changes in precipitation before and after the operations. To ensure the diversity of the training set, the training samples of this invention cover the following aspects:

[0079] Different meteorological conditions: The training samples cover rain enhancement operations under different meteorological backgrounds, such as convective precipitation and stratiform precipitation, so that the model can learn the rain enhancement effects under different weather conditions.

[0080] Different types of operations: The samples include different types of artificial rainmaking operations, such as rockets, airplanes, and anti-aircraft guns, so that the model can learn the impact of different operations on precipitation.

[0081] Different seasons and regions: The effectiveness of rain enhancement operations is affected by season and region. Therefore, the training samples need to include operational data from different seasons and geographical areas to ensure that the model can cope with various environmental conditions.

[0082] By integrating these sampling strategies and training set construction methods, the model of this invention can learn the impact of rain enhancement operations on precipitation under different conditions, thereby improving the accuracy of prediction.

[0083] Step 103: Generate a first tag based on the actual precipitation, and generate a second tag based on the actual precipitation and the precipitation without rain enhancement operations.

[0084] The actual precipitation is used as the first label; the second label is generated based on the difference between the actual precipitation and the precipitation without rain enhancement operations.

[0085] First Label: The first label data primarily comes from CMPAS precipitation data, representing the actual precipitation after the rain enhancement operation. For each time step, the precipitation data provided by CMPAS, in mm / h or mm, is used as a supervision signal to construct the target precipitation tensor. The label tensor has a shape of [T, H, W], consistent with the spatial dimensions of the input data. This first label is used for training the main model.

[0086] Step 104: Based on the input tensor and the first label, perform model training and calculate the first loss function.

[0087] Among them, L MSE The first loss function is calculated as follows:

[0088]

[0089] Among them, P pred It refers to the predicted precipitation amount; that is, after each model training, a predicted precipitation result is obtained, P. true This indicates the amount of precipitation without rain enhancement operations.

[0090] Specifically, the spatiotemporal model used is a Convolutional Long Short-Term Memory (ConvLSTM) network, which is a spatiotemporal model combining Convolutional Neural Networks (CNN) and LSTM, suitable for processing spatiotemporal sequence data. The ConvLSTM network can effectively extract spatiotemporal features and identify the complex relationship between meteorological fields such as temperature, humidity, and wind speed and precipitation. Its main task is to predict precipitation based on input meteorological data, and it follows physical constraints such as mass conservation and momentum conservation during training.

[0091] ConvLSTM is an extension based on LSTM. It uses convolution operations instead of the fully connected operations in standard LSTM, enabling the network to preserve spatial structure while effectively capturing dynamic changes in time series data when processing spatiotemporal data. The main advantage of ConvLSTM is that the convolution operation allows the network to share weights spatially, enabling efficient processing of high-dimensional data with spatial dependencies, such as weather fields.

[0092] In the model of this application, ConvLSTM first extracts meteorological features through spatiotemporal convolution operations, and then uses LSTM units to capture time dependencies to generate precipitation prediction results. However, the simple ConvLSTM model still cannot accurately quantify the precipitation increment ΔP caused by rain enhancement operations. Therefore, it is necessary to introduce operation perturbation guidance and physical constraints, which will be further detailed below.

[0093] The core of ConvLSTM lies in the combination of convolution operations and LSTM gating mechanisms. Specifically, ConvLSTM consists of the following main parts:

[0094] Input gate (i t ): Determines the degree to which the current input information updates the cell state.

[0095] Forgotten Gate (f t ): Determines the proportion of the cell state from the previous time step that is retained in the current cell state.

[0096] Output gate (o) t ): Determines the degree to which the cell state affects the current output.

[0097] Cell state (C t ): Carries long-term memories about time series.

[0098] Hidden state (H) t ): Output status, including spatial information of the current time step.

[0099] The calculation formula for each ConvLSTM layer is explicitly expressed as follows:

[0100] i t =σ(W xi *X t +W hi *H t-1 +b i )

[0101] f t =σ(W xf *X t +W hf *H t-1 +b f )

[0102] o t =σ(W xo *X t +W ho *H t-1 +b o )

[0103]

[0104] Among them, X t This is the input at the current moment, typically meteorological data (such as temperature, humidity, wind speed, etc.). H t W represents the hidden state at the current time step, containing spatial feature information of that moment. σ is the sigmoid activation function, and tanh is the hyperbolic tangent activation function. xi W hi W xf W hf W xo W ho W xc W hc These are the kernel weights, and * indicates the convolution operation. This indicates element-wise multiplication.

[0105] The convolution operation in each gate allows the network to share weights within local regions, significantly improving computational efficiency when processing high-dimensional meteorological data. Cell state C t It stores long-term dependency information from the time series, while the hidden state H t This is used to output the spatial characteristics at the current moment.

[0106] The ConvLSTM backbone network extracts spatiotemporal features by stacking multiple convolutional layers and LSTM units. The ConvLSTM layers in the model process the spatial and temporal information of the input data respectively, capturing local spatial features through convolutional operations and capturing the dynamic changes of the time series through the LSTM gating mechanism.

[0107] Each ConvLSTM unit processes the input data through convolution operations, progressively extracting spatiotemporal features. The network consists of three ConvLSTM layers, with the number of output channels gradually increasing in each layer to enhance the network's representational power.

[0108] The first layer of ConvLSTM typically uses a small convolutional kernel, such as 3×3, and a small number of channels, such as 32 channels, and is responsible for extracting low-level spatial features from the input data.

[0109] The second layer, ConvLSTM, increases the kernel size, such as 5×5, and the number of channels, such as 64 channels, to capture higher-order spatiotemporal patterns.

[0110] The third ConvLSTM layer is the output layer. This layer decodes the hidden state, mapping the information to the precipitation intensity prediction space. This layer outputs 32 channels, restoring the spatial resolution through convolution operations and generating a precipitation prediction map of the same size as the input meteorological data.

[0111] The output layer of this model is:

[0112] Finally, a 1×1 convolutional layer maps the output of the ConvLSTM to the precipitation prediction space. This layer outputs a precipitation prediction map with the same spatial resolution as the input data, representing the precipitation distribution at future times.

[0113] The modeling process involves three modules, including:

[0114] First, the disturbance guidance module

[0115] The perturbation guidance module is one of the core components of this invention. Its design aims to guide relevant information about rain enhancement operations, such as operation type, intensity, and timing, into the input of the ConvLSTM model, enabling the model to focus on the operation's impact on precipitation. By incorporating operation perturbation information into the model's input data, the model can learn the precipitation change patterns caused by rain enhancement operations.

[0116] Therefore, it is necessary to embed the perturbation information into the input tensor in a way that has physical and spatial meaning, so that it can play a guiding role at time step t and spatial grid point (i,j). The construction of the input tensor described above in this application can be directly input into this perturbation guidance module, which corresponds to the first loss function.

[0117] Second, the construction of the physical constraint module.

[0118] The construction of the physical constraint module corresponds to step 105 of this application.

[0119] Step 105: Calculate the second loss function based on the input tensor and the preset physical constraint model.

[0120] In this application, the preset physical constraint model refers to the physical constraint module. In the construction of this physical constraint module, information such as water vapor content, temperature, and specific heat capacity are included in the meteorological information. This information can be directly collected by sensors, or indirectly collected by performing some calculations on the parameters collected by sensors. As for how to collect this information, this is existing technology in the field and will not be described in detail here.

[0121] according to Determine the mass conservation constraints; where q v,t q represents the water vapor content at the current time t. v，t+Δt P represents the water vapor content at the next time step t+Δt. t+Δt H represents the precipitation at the next time step t+Δt, H represents the height of the spatial grid, and W represents the width of the spatial grid.

[0122] according to Determine the thermodynamic conservation constraints; where cp represents the specific heat capacity of air, ΔT represents the temperature change, Lv represents the latent heat of water vapor condensation, and P represents the precipitation.

[0123] according to Determine the momentum conservation constraint term; where v t Let v represent the wind speed field at the current time t. t+Δt Let ρ represent the wind speed field at the next time step t+Δt, and let ρ represent the air density. represents the pressure gradient, and g represents the acceleration due to gravity;

[0124] According to L phys =λ1L mass +λ2L thermo +λ3L momentum Determine the physical constraint loss term, where L phy L represents the physical constraint loss term. mass L thermoL momentum These are the loss terms for the constraints of mass conservation, thermodynamic conservation, and momentum conservation, respectively. λi is the weighting coefficient of the constraint term, and the physical constraint loss term is the second loss function.

[0125] The physical constraints in this application ensure that the model's output conforms to actual physical laws. For example, constraints such as mass conservation, thermodynamic conservation, and momentum conservation force the model to follow fundamental laws of atmospheric physics. These constraints not only improve the model's physical consistency but also enhance its ability to explain actual meteorological phenomena. The construction of the second loss function is explained in detail below.

[0126] 1) Mass conservation constraint

[0127] The mass conservation constraint ensures the balance between atmospheric water vapor and precipitation. In this model, this constraint is achieved by calculating the difference between the increase in precipitation and the decrease in water vapor content. Specifically, the difference between the atmospheric water vapor content at the initial and final times and the difference between the precipitation are brought close to zero, i.e.:

[0128]

[0129] Where qv,t represents the water vapor content at the current time t, in kg / kg, which can be obtained from the variable specific humidity output by CMA-MESO;

[0130] q v，t+Δt This represents the water vapor content at the next time step t+Δt, in kg / kg, also derived from CMA-MESO output data.

[0131] P t+Δt The precipitation at the next time step t+Δt, in mm / h, can be obtained directly from the precipitation forecast results of CMA-MESO.

[0132] H and W represent the height and width of the model space grid, respectively, indicating the dimension of the space grid.

[0133] This constraint ensures the balance between the decrease in water vapor content and the increase in precipitation at each time step, thus constraining the balance between precipitation and water vapor and forcing the model's predictions to satisfy the law of conservation of mass.

[0134] 2) Thermodynamic conservation constraints

[0135] Thermodynamic conservation constraints are used to ensure energy conservation during precipitation. The condensation of water vapor in the atmosphere releases latent heat and affects air temperature; therefore, the model needs to ensure that the relationship between predicted precipitation and temperature changes is reasonable. This constraint is calculated using the following formula:

[0136]

[0137] Where cp is the specific heat capacity of air (unit: J / kg·K), which is a constant value;

[0138] ΔT is the temperature change, measured in K, and can be calculated using air temperature data provided by CMA-MESO. It represents the temperature change before and after precipitation.

[0139] Lv is the latent heat of water vapor condensation, with units of J / kg. It is usually set to 2.5 × 10⁶ J / kg, based on atmospheric physical standard values.

[0140] P represents precipitation, measured in mm / h, and is obtained directly from CMA-MESO output data.

[0141] This constraint ensures that the increase in precipitation does not exceed thermodynamic limits and matches the latent heat released by water vapor condensation.

[0142] 3) Momentum conservation constraint

[0143] Momentum conservation constraints are used to ensure the reasonableness of the dynamic relationship between wind field and precipitation predicted by the model. In weather simulation, changes in wind field are closely related to precipitation generation. Through the momentum equation, the model's predictions must satisfy the law of conservation of momentum:

[0144]

[0145] Among them, v t This is the wind speed field at the current time t, in m / s, which can be obtained from the CMA-MESO wind speed field data.

[0146] v t+Δt It is the wind speed field at the next moment t+Δt;

[0147] ρ is the density of air, with units of kg / m³. 3 It is usually a constant that depends on temperature and air pressure;

[0148] It is the pressure gradient, with units of Pa / m, which can be calculated using CMA-MESO pressure field data;

[0149] g is the acceleration due to gravity, measured in m / s². 2 Take a constant value of 9.81 m / s 2 This constraint ensures that changes in precipitation are consistent with the conservation of momentum.

[0150] 4) Construction of the total loss function for physical constraints

[0151] The total physical constraint loss function, also known as the second loss function, integrates the constraints of mass conservation, thermodynamic conservation, and momentum conservation. The formula is as follows:

[0152] L phys =λ1L mass +λ2t hermo +λ3L momentum

[0153] L phy L represents the physical constraint loss term, also known as the second loss function. mass L thermo L momentum These are the loss terms for the constraints of mass conservation, thermodynamic conservation, and momentum conservation, respectively, and λi is the weighting coefficient of the constraint term.

[0154] Third, the construction of the causal inference module

[0155] The core task of the causal inference module is to identify and quantify the direct impact of artificial rain enhancement operations on precipitation. Unlike traditional statistical models that only establish the correlation between input and output, causal inference emphasizes the interventionist question of "whether a similar precipitation response would have occurred if a certain operation had not been carried out." Essentially, it quantifies, infers, and predicts the causal effect between "operation-precipitation." In this invention, this module serves as the backbone model, an extension of ConvLSTM. Based on the trained operation perturbation response features, and with the aid of physical constraints, it infers the precipitation increment ΔP caused by each artificial operation, forming a complete "operation-perturbation-response" closed-loop path.

[0156] To effectively extrapolate the causal effects of artificial rain enhancement operations on precipitation response, this invention preferably employs the T-Learner model, implemented using the EconML framework, as the core method for the causal extrapolation module. T-Learner (Two-model Learner) is a classic strategy for estimating heterogeneous causal effects, suitable for scenarios with highly variable spatial scales and meteorological conditions. This method derives the response difference, ΔP, caused by artificial intervention by separately modeling "precipitation prediction under operational conditions" and "precipitation prediction without operational conditions." It boasts advantages such as clear structure, strong scalability, and stable training, making it particularly suitable for the operational disturbance intervention scenario constructed using CMPAS and CMA-MESO data in this invention.

[0157] Since CMA-MESO does not consider human intervention disturbances, Ymeso can approximate the "natural field" precipitation when no operations are carried out; while CMPAS integrates objective precipitation observations from actual measurements, including responses that may be caused by human interventions. Therefore, the difference between the two, ΔPref, can be regarded as the anthropogenic precipitation enhancement effect under the current meteorological background and operational conditions, i.e., the target variable that the model needs to learn.

[0158] This loss guides the model to explicitly learn the response difference caused by the task behavior, ensuring that the model output has quantifiable causal significance, which is the core of the causal structure of this invention.

[0159] The construction of the causal deduction module corresponds to steps 106 to 110 of this application.

[0160] Step 106: Based on the input tensor and the second label, determine the first precipitation forecast during the operation and the second precipitation forecast when there is no rain enhancement operation.

[0161] Specifically, during the training phase, the input tensor is divided into two groups based on the input tensor, constructing two data subsets S1 = {(Xi,Yi)|Ti = 1} and S0 = {(Xi,Yi)|Ti = 0}; where Xi in (Xi,Yi) represents the input tensor of the i-th sample, Yi represents the actual precipitation of the i-th sample; Ti = 1 represents the time when artificial rain enhancement operations are conducted, and Ti = 0 represents the time when artificial rain enhancement operations are not conducted.

[0162] Two sub-models, f1(X) and f0(X), are trained using two subsets of data. f1(X) represents the precipitation response when artificial rain enhancement is in operation and is trained using S1, while f0(X) represents the precipitation response when artificial rain enhancement is not in operation and is trained using S0.

[0163] Based on f1(X) from the two trained sub-models, the first precipitation prediction is obtained;

[0164] The second precipitation prediction is obtained based on f0(X) from the two trained sub-models.

[0165] Step 107: Calculate the first loss term of the first precipitation forecast and the second loss term of the second precipitation forecast;

[0166] Specifically, according to Construct the operation disturbance sample loss, and represent the first loss term through the operation disturbance sample loss, where f1(X) represents the precipitation response when the predicted operation disturbance exists, N1 represents the number of operation samples, and Yi is the actual precipitation in the fused precipitation information;

[0167] according to Construct a non-operational disturbance sample loss, and use the non-operational disturbance sample loss to represent the second loss term, where f0(X) represents the natural precipitation response when the predicted operational disturbance does not exist, and N0 represents the number of non-operational samples.

[0168] Step 108: Determine the first difference based on the first precipitation forecast and the second precipitation forecast;

[0169] Where ΔP = f1(X) - f0(X) represents the first difference.

[0170] Step 109: Determine the causal supervision loss term based on the first difference and the second label;

[0171] Specifically, based on the first difference and the second label, through Determine the causal monitoring loss item;

[0172] Where ΔPi represents the first difference of the i-th sample, ΔPrefi represents the second label of the i-th sample; ΔPref = Yobs – Ymeso represents the second label; Yobs represents the actual precipitation; Ymeso represents the precipitation without rain enhancement operations, and ΔPref represents the second label, which is used to represent the response magnitude caused by the disturbance.

[0173] Step 110: Determine the third loss function based on the first loss term, the second loss term, and the causal supervision loss term.

[0174] Specifically, through Calculate the causal inference loss term; where α, β, and γ are the principal loss balance coefficients. These are the loss for operational disturbance samples, the loss for non-operational disturbance samples, and the loss for causal response monitoring, respectively.

[0175] This causal inference module ensures that the model has high prediction accuracy under both operational and non-operational conditions, and its output difference has causal explanatory power, which is the key contribution of this invention to the intelligent evaluation of artificial rain enhancement.

[0176] Step 111: Optimize the model based on the first loss function, the second loss function, and the third loss function until the model converges;

[0177] Step 112: When the model converges, output the precipitation increment of the rain enhancement operation.

[0178] Specifically, in the complete training framework, the model's total loss function consists of three parts:

[0179] Main task loss: measures the error between the model's predicted precipitation and the actual precipitation, i.e., the first loss function;

[0180] Physical constraint loss: ensures that the model output conforms to physical laws, i.e., the second loss function;

[0181] Causal inference loss: measures the difference between the model-inferred ΔP and the reference value ΔPrefi, i.e., the third loss function.

[0182] During training, these three loss terms will be optimized simultaneously by setting appropriate weight coefficients to balance their importance. For example, when the focus is on prediction accuracy in the early stages of model training, the weight of the main task loss can be increased; while in the later fine-tuning stage, the proportion of causal loss and physical constraint loss can be appropriately increased to improve the causal interpretability and scientific validity of the model.

[0183] The total loss function integrates the main task, physical constraints, and causal inference, and the formula is as follows:

[0184] L total =L MSE +λ1L phys +λ2L causal

[0185] L MSE The error in the main model's precipitation prediction for Yobs is represented by the conventional precipitation regression loss, which measures the difference between the model's predicted precipitation and the actual precipitation.

[0186] L phys L represents the total loss due to physical constraints. causal λi represents the total loss for causal inference, and λi is the weight coefficient of the constraint term, which controls the optimization balance among multiple tasks.

[0187] This loss function guides the model to consider both prediction accuracy and the satisfaction of physical constraints during the optimization process, thereby improving the physical consistency of the model.

[0188] L MSE The first loss function is calculated as follows:

[0189]

[0190] Among them, P pred It is the precipitation predicted by the model, P truo This is the actual amount of rainfall.

[0191] The final trained model can not only predict precipitation under different meteorological conditions and operational disturbances, but also output the "additional precipitation" ΔP brought about by the operation, truly realizing precipitation response simulation based on operational disturbances.

[0192] When training deep neural networks, especially ConvLSTM models processing spatiotemporal sequence data, the choice of optimization algorithm has a crucial impact on the model's convergence speed and final performance. This invention uses a combination of gradient clipping and the AdamW optimizer to improve the model's training efficiency. This method combines the stability of gradient clipping with the adaptive learning rate adjustment advantages of AdamW, effectively addressing gradient explosion and overfitting problems in long sequence training. The following section discusses in detail the principles, implementation details, and considerations for practical applications of this method.

[0193] The core idea of ​​gradient clipping is to prevent gradient explosion by limiting the magnitude of the gradient, thereby ensuring the stability of the training process. In deep networks, especially models with recurrent structures like ConvLSTM, gradients may accumulate between time steps during backpropagation, leading to excessively large gradient magnitudes. This phenomenon is particularly common in long-sequence tasks such as meteorological data modeling. This invention uses norm-based clipping to achieve gradient clipping, which is more conducive to the stability of the optimization process. The specific formula is as follows:

[0194]

[0195] Where grad is the gradient of the current parameter, which is a tensor, t is a preset threshold, such as 1.0, and ||grad|| is the L2 norm of the gradient.

[0196] Specifically, when the L2 norm of the gradient exceeds a preset threshold, the gradient will be scaled down proportionally so that its norm equals the threshold.

[0197] The AdamW optimizer is an improved version of the Adam algorithm. It resolves the conflict between L2 regularization and adaptive learning rate in the traditional Adam optimizer by decoupling the weight decay and gradient update steps. In the standard Adam optimizer, the weight decay term is directly added to the gradient, which can cause the adaptive learning rate to interfere with the effect of weight decay. AdamW, however, applies weight decay independently of gradient updates, directly to the parameters themselves, thus more effectively controlling model complexity and preventing overfitting. In AdamW's update formula, the momentum calculation and bias correction steps remain consistent with standard Adam, but the weight decay term is handled separately during parameter updates. This improvement allows AdamW to maintain the advantages of adaptive learning rate while performing more stable regularization, resulting in better generalization performance in tasks such as image classification, natural language processing, and spatiotemporal sequence prediction.

[0198] AdamW's update process consists of the following steps:

[0199] (1) Calculate the first and second momentum.

[0200] m t =β1·m t-1 +(1-β1)·g t

[0201]

[0202] Among them, g t This represents the gradient at the current time step, i.e., the clipped gradient; m t The first-order momentum, i.e., the mean gradient; v t β1 and β2 are the second-order momentum, i.e., the gradient variance; β1 and β2 are the momentum decay rates, with default values ​​of 0.9 and 0.999.

[0203] (2) Deviation correction

[0204]

[0205] Where t is the current training step number, starting from 1; m t and v t These are the corrected momentum and variance, respectively, to address the initial zero bias problem;

[0206] (3) Parameter update

[0207]

[0208] Where, θ t The updated model parameters are η, which is the learning rate, such as 3e-4; ∈ is the numerical stability term, which defaults to 1e-8; and λ is the weight decay coefficient, such as 0.01.

[0209] By properly setting the parameters in the above formula, gradient clipping + AdamW can efficiently and stably train ConvLSTM models, especially performing well in long sequence spatiotemporal prediction tasks. In practical applications, the threshold and learning rate need to be adjusted according to the task.

[0210] Training process

[0211] During training, the model's parameters are updated by minimizing the total loss function. This loss function considers not only traditional prediction errors but also physical constraints. The specific training process can be divided into the following steps:

[0212] 1) Data preparation and input

[0213] (1) Obtain meteorological data from CMA-MESO, including temperature, humidity, wind speed, precipitation, etc.

[0214] (2) Obtain real precipitation data from CMPAS.

[0215] (3) After being processed by the perturbation guidance module, the input data is fed into the ConvLSTM network, which extracts spatiotemporal features through convolution operations. The perturbation guidance module ensures that information on artificial rain enhancement operations is fully reflected in the input, thereby guiding the model to focus on the areas affected by the operations on precipitation.

[0216] 2) Forward propagation

[0217] At each time step, the ConvLSTM network calculates the spatiotemporal characteristics of the meteorological field through forward propagation and generates precipitation predictions. These predictions are compared with actual precipitation data provided by CMPAS to calculate prediction error losses, such as mean squared error (MSE). Furthermore, the network calculates physical constraints, such as mass conservation, thermodynamic conservation, momentum conservation, and cloud microphysical conservation. By comprehensively processing these constraints, the network ensures that the model's output conforms to physical laws.

[0218] 3) Physical constraint calculation

[0219] In each forward propagation, physical constraints are calculated based on the model output and CMA-MESO meteorological data. These physical constraints include mass conservation, thermodynamic conservation, momentum conservation, and cloud microphysical conservation. Each physical constraint, along with the prediction error loss, participates in the calculation of the total loss. The physical constraints are added to the total loss function through a weighted summation with the prediction error loss, thereby guiding the model's learning and ensuring its physical consistency on a data-driven basis.

[0220] 4) Backpropagation and gradient update

[0221] Backpropagation is a crucial step in training. In this stage, gradient clipping is used to prevent gradient explosion. Backpropagation is performed by calculating the gradient of the loss function relative to the model parameters to obtain the gradient for each model parameter. To prevent excessively large gradients from causing training instability, gradient clipping is applied, which limits the maximum value of the gradient to ensure stability during the update process.

[0222] Gradient clipping is implemented through the following steps:

[0223] (1) Calculate the L2 norm of all gradients.

[0224] (2) If the L2 norm of the gradient exceeds the preset threshold, scale all gradients proportionally until their L2 norm is equal to the set maximum value.

[0225] (3) Backpropagation includes the gradient effect of physical constraints. Each physical constraint (such as mass conservation, thermodynamic conservation, etc.) will correct the corresponding gradient, thereby ensuring that the model always follows physical laws during the learning process.

[0226] After calculating the gradient, the AdamW optimizer is used to update the parameters. Compared with the traditional Adam optimizer, AdamW regularizes the weights through a weight decay mechanism, thereby effectively preventing overfitting and improving the model's generalization ability.

[0227] By combining gradient clipping with the AdamW optimizer, the model training becomes more stable and effectively prevents gradient explosion and overfitting problems.

[0228] 5) Iterative training and convergence

[0229] Through multiple iterations, the model continuously adjusts its parameters to minimize the total loss function, thereby optimizing both prediction error and physical constraints. Each update is based on the gradient information from the previous iteration, thus the model's performance gradually improves, eventually converging after a sufficient number of training epochs.

[0230] Second, model output

[0231] After training, the model will perform real-time evaluation and output the precipitation increment (ΔP) caused by the operation. This increment value can be compared with CMPAS measured precipitation data to verify the accuracy of the model's prediction results. In this way, the model can provide meteorological departments with real-time and accurate evaluation of the effectiveness of artificial rain enhancement operations, assisting in decision-making and optimizing rain enhancement strategies.

[0232] By inputting meteorological information, operational information, and fused precipitation information at any given time into a pre-trained model, the precipitation output can be predicted, thereby enabling the prediction of precipitation increments from rain enhancement operations.

[0233] By applying the precipitation response extrapolation method based on physical constraints and operational perturbation provided in this invention, a spatiotemporal feature extraction network based on ConvLSTM is used to capture spatial correlations through convolutional operations and to model temporal dependencies using the LSTM mechanism, achieving a deep understanding of atmospheric evolution. The physical constraint module introduces fundamental meteorological principles such as mass conservation, energy balance, and momentum constraints as hard constraints to ensure that the model output conforms to physical laws. The causal extrapolation framework accurately separates the effects of operational perturbation and natural precipitation through parallel training of intervention and control branches, ultimately outputting a precipitation increment ΔP with causal explanation. In terms of model optimization, a multi-task joint loss function is adopted, comprehensively considering prediction accuracy, regression loss, causal consistency, causal response loss, physical rationality, and constraint loss. Furthermore, the AdamW optimizer and gradient pruning technique are innovatively combined to effectively improve the stability and convergence speed of model training.

[0234] Compared to traditional methods, this invention offers three significant advantages: First, through spatiotemporal dynamic modeling and precise perturbation coding, it achieves a refined assessment of the impact of artificial rain enhancement operations; second, the introduction of physical constraint mechanisms ensures the meteorological rationality of the forecast results; and finally, the causal inference framework provides interpretable quantitative indicators of operational effectiveness. This method offers a real-time, accurate, and interpretable operational effectiveness evaluation mechanism, overcoming the limitations of traditional methods that rely on simulation and speculation. It can provide reliable decision support for meteorological departments and optimize rain enhancement operation strategies. By comprehensively utilizing meteorological forecast data, operational information, causal inference, and physical constraints, this invention has broad application prospects in the field of artificial rain enhancement operation evaluation, particularly in disaster response, environmental protection, and meteorological intervention.

[0235] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0236] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0237] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A precipitation response extrapolation method based on physical constraints and operational disturbance guidance, characterized in that, The method includes: According to a preset sampling strategy, meteorological information, surface-fused precipitation information, and artificial rain enhancement operation information are acquired; the meteorological information includes precipitation before rain enhancement operations; the surface-fused precipitation information includes the actual precipitation during rain enhancement operations. Based on the meteorological information and the operational information, an input tensor is generated; the input tensor includes a time dimension, a spatial dimension, the number of meteorological element channels, and a disturbance channel; A first tag is generated based on the actual precipitation, and a second tag is generated based on the actual precipitation and the precipitation without rain enhancement operations. Based on the input tensor and the first label, a convolutional long short-term memory network model is trained, and a first loss function is calculated. The second loss function is calculated based on the input tensor and the preset physical constraint model; Based on the input tensor and the second label, the first precipitation forecast during the operation and the second precipitation forecast without rain enhancement operation are determined respectively. Calculate the first loss term for the first precipitation forecast, and calculate the second loss term for the second precipitation forecast; Based on the first precipitation forecast and the second precipitation forecast, determine the first difference; Based on the first difference and the second label, determine the causal supervision loss term; Based on the first loss term, the second loss term, and the causal supervision loss term, a third loss function is determined; Based on the first loss function, the second loss function and the third loss function, the convolutional long short-term memory network model is optimized until the convolutional long short-term memory network model converges. When the convolutional long short-term memory network model converges, it outputs the precipitation increase from the rain enhancement operation. The step of training a convolutional long short-term memory network model based on the input tensor and the first label, and calculating the first loss function specifically includes: according to Calculate the main task loss term, which is the first loss function; where P pred This is the precipitation predicted by the convolutional long short-term memory network model, P. true This refers to the rainfall without rain enhancement operations; The calculation of the second loss function based on the input tensor and the preset physical constraint model specifically includes: according to Determine the mass conservation constraints; where q v,t q represents the water vapor content at the current time t. v，t+Δt P represents the water vapor content at the next time step t+Δt. t+Δt H represents the precipitation at the next time step t+Δt, W represents the height of the spatial grid, and x and y represent the spatial coordinate variables. according to Determine the thermodynamic conservation constraints; where c p L represents the specific heat capacity of air, ΔT represents the temperature change. v P represents the latent heat of water vapor condensation, and P represents precipitation. according to Determine the momentum conservation constraint term; where V t V represents the wind speed field at the current time t. t+Δt Let ρ represent the wind speed field at the next time step t+Δt, and let ρ represent the air density. represents the pressure gradient, and g represents the acceleration due to gravity; This represents the derivative of the wind speed field v with respect to itself; According to L phys =λ1L mass +λ2L thermo +λ3L momentum Determine the physical constraint loss term, where L phy L represents the physical constraint loss term. mass L thermo L momentum These are the loss terms for the constraints of mass conservation, thermodynamic conservation, and momentum conservation, respectively. λ1, λ2, and λ3 are the weighting coefficients for the constraints of mass conservation, thermodynamic conservation, and momentum conservation, respectively. The physical constraint loss term is the second loss function. The step of determining the third loss function based on the first loss term, the second loss term, and the causal supervision loss term specifically includes: pass Calculate the causal inference loss term; where α, β, and γ are the principal loss balance coefficients. These are the loss for operational disturbance samples, the loss for non-operational disturbance samples, and the causal supervision loss term, respectively.

2. The method according to claim 1, characterized in that, The step of generating the input tensor based on the meteorological information and the operational information specifically includes: The meteorological information includes multiple meteorological variables, through... The meteorological variables are standardized; where X is the meteorological variable, μ is the mean of the meteorological variable, and σ is the standard deviation of the meteorological variable. The standardized meteorological information is processed into a four-dimensional tensor [T,H,W,C]. meteo ], where T represents the time step, H represents the height of the spatial grid, W represents the width of the spatial grid, and C meteo The number of channels representing meteorological variables; Align the time of the operation information with the timestamps of the meteorological variables; Map the time range from the start time to the end time of the task information to the effective time step in the perturbation tensor; Convert the dosage or intensity of the application in the operation information into perturbation values; according to Construct a perturbation field and represent the operation perturbation channel using a spatial mask or perturbation vector to generate the operation perturbation tensor [T,H,W,1]; where D t0 Let S represent the perturbation field, and let S represent the perturbation intensity value, where S∈[0,1]. Based on the four-dimensional tensor and the operation perturbation tensor, generate the input tensor [T,H,W,C]. meteo +1], where C meteo It represents the channel number of the meteorological variable, with +1 indicating the disturbance channel.

3. The method according to claim 1, characterized in that, The preset sampling strategy includes event-driven sampling and control sample sampling. Event-driven sampling uses the moment when artificial rain enhancement operation occurs as the center of the sampling window and extracts data before and after the artificial rain enhancement operation according to a preset step size. Control sample sampling collects data when there is no artificial rain enhancement operation.

4. The method according to claim 1, characterized in that, The process of generating a first tag based on the actual rainfall and a second tag based on the actual rainfall and the rainfall without rain enhancement operations specifically includes: Use actual precipitation as the primary label; A second label is generated based on the difference between the actual precipitation and the precipitation without rain enhancement operations.

5. The method according to claim 1, characterized in that, The step of determining the first precipitation forecast during the operation and the second precipitation forecast without rain enhancement operation based on the input tensor and the second label specifically includes: During the training phase, the input tensor is divided into two groups based on the input tensor, constructing two data subsets S1 = {(Xi,Yi)|Ti = 1} and S0 = {(Xi,Yi)|Ti = 0}; where Xi in (Xi,Yi) represents the input tensor of the i-th sample, and Yi represents the actual precipitation of the i-th sample; Ti = 1 represents the time when artificial rain enhancement operations are conducted, and Ti = 0 represents the time when artificial rain enhancement operations are not conducted; Two sub-models, f1(X) and f0(X), are trained using two subsets of data. f1(X) represents the precipitation response when artificial rain enhancement operations are in place and is trained using S1, while f0(X) represents the precipitation response when artificial rain enhancement operations are not in place and is trained using S0. Based on f1(X) from the two trained sub-models, the first precipitation prediction is obtained; The second precipitation prediction is obtained based on f0(X) from the two trained sub-models.

6. The method according to claim 1, characterized in that, The calculation of the first loss term for the first precipitation forecast and the calculation of the second loss term for the second precipitation forecast specifically include: according to Construct the operation disturbance sample loss, and represent the first loss term through the operation disturbance sample loss, where f1(X) represents the precipitation response when artificial rain enhancement operation is present, N1 represents the number of operation samples, and Yi represents the actual precipitation of the i-th sample; according to Construct the non-operational disturbance sample loss, and use the non-operational disturbance sample loss to represent the second loss term, where f0(X) represents the precipitation response when there is no artificial rain enhancement operation, and N0 represents the number of non-operational samples.

7. The method according to claim 1, characterized in that, The step of determining the causal supervision loss term based on the first difference and the second label specifically includes: Based on the first difference and the second label, through Determine the causal monitoring loss item; Where ΔPi represents the first difference of the i-th sample, ΔPrefi represents the second label of the i-th sample; ΔPref=Yobs–Ymeso represents the second label; Yobs represents the actual precipitation; Ymeso represents the precipitation without rain enhancement operations; the second label represents the response magnitude caused by the disturbance; ΔP = f1(X) - f0(X) represents the first difference. The term represents the causal monitoring loss term, f1(X) represents the precipitation response when artificial rain enhancement operations are in place, f0(X) represents the precipitation response when artificial rain enhancement operations are not in place, and N represents the sum of the number of operational samples and the number of non-operational samples.

Images