A short-time precipitation nowcasting method fusing quantum computing and deep learning model

Abstract

The application provides a short-time precipitation nowcasting method fusing quantum computation and a deep learning model, and belongs to the field of short-time precipitation nowcasting, and steps are as follows: firstly, long-term historical data of precipitation and meteorological elements of a target area are extracted; then, the precipitation at a time of tentative prediction is taken as a prediction target quantity, and meteorological elements and precipitation of M hours before the time of tentative prediction are taken as model input characteristic quantities; then, a quantum computation method is used to change the meteorological elements and precipitation of M hours before the time of tentative prediction into high-dimensional quantum probability characteristics about the precipitation at the time of tentative prediction; finally, quantum circuit precipitation probability information is taken as input, and the precipitation of 1 hour in the future is taken as output, a short-time precipitation nowcasting model is constructed and trained, and the application uses quantum circuits and quantum superposition states to map classical data to a high-dimensional quantum Hilbert space, can avoid a local optimal solution trap, significantly reduces a calculation burden, and realizes short-time precipitation prediction under the condition of massive high-dimensional meteorological element inputs.

Description

Technical Field

[0001] This invention belongs to the field of short-term precipitation nowcasting, specifically involving a short-term precipitation nowcasting method that integrates quantum computing and deep learning models. Background Technology

[0002] In meteorological operations, short-duration heavy precipitation refers to precipitation with a minimum of 20 mm per hour, while extreme short-duration heavy precipitation is defined as precipitation with a minimum of 50 mm per hour. This phenomenon is characterized by its suddenness, localization, and short duration, making it difficult to forecast, especially when the heavy precipitation is scattered and short-lived.

[0003] Regarding the current state of research, short-duration heavy precipitation forecasting mainly falls into two categories: observational methods and simulation methods. Observational methods primarily rely on multi-source data fusion, such as Global Navigation Satellite System (GNSS) technology for detecting water vapor distribution. Due to its real-time, all-weather, and wide-area coverage advantages, it has become an important tool for analyzing short-duration heavy precipitation. However, the spatial resolution of current observational equipment (such as GNSS and radar) is still not fine enough, making it difficult to capture the minute-scale characteristics of short-duration precipitation, especially in complex terrain or urban areas, where the high degree of unevenness in precipitation distribution can lead to missed or false alarms. Furthermore, in practical operations, there are still time delays ranging from several minutes to tens of minutes in the acquisition, transmission, quality control, and fusion processing of observational data, which also poses a challenge to the timeliness of nowcasting of short-duration heavy precipitation based on observational methods. In addition, the accuracy of observational data is also affected by factors such as equipment status and environmental interference (such as electromagnetic signal interference and cloud cover). Especially under extreme weather conditions that lead to short-duration heavy precipitation, missing data or outliers may further reduce the reliability of the forecast.

[0004] Simulation and forecasting methods are mainly divided into two categories: mesoscale weather model forecasting methods and data-driven model forecasting methods. Mesoscale weather models simulate and predict precipitation by solving partial differential equations that combine precipitation with dynamic and thermodynamic physical quantities. However, this method also has many limitations. For example, mesoscale models are highly sensitive to initial fields and boundary conditions; small errors can be amplified during the simulation, affecting the accuracy of precipitation intensity and location forecasts. Furthermore, precipitation calculations in mesoscale models involve complex microphysical, convection, and boundary layer parameterization schemes; different scheme choices can lead to significant differences in forecast results, increasing forecast uncertainty. While data assimilation techniques in mesoscale models (such as EnKF and 3D / 4D-Var methods) can improve forecast accuracy, their high technical barriers, large computational load, and stringent requirements on the quality and quantity of observational data make them difficult to popularize and continuously update in all operational scenarios. Furthermore, high-resolution simulations of mesoscale models require enormous computing resources and storage space, making it difficult to achieve rapid iteration and real-time updates in operational use, thus limiting their application in short-term nowcasting.

[0005] Data-driven models primarily utilize machine learning or deep learning models to automatically extract precipitation evolution features from large amounts of historical observation data, achieving high-precision and high-efficiency nowcasting. Compared to mesoscale models, data-driven models are typically more accurate and computationally efficient for precipitation forecasts within 2 hours. However, data-driven models usually rely on high-quality, large-sample meteorological labeled data to train complex machine learning or deep learning models. These massive historical meteorological training datasets with high feature dimensions, and the complex machine learning or deep learning models with large parameter counts, all contribute to a heavy computational and optimization burden, slowing down the model training process and thus affecting the timeliness of short-term heavy precipitation nowcasting. Furthermore, as the amount of historical training data increases and the data dimensionality rises, the dimensionality of its state space grows exponentially, and classical computers encounter computational bottlenecks when handling this complexity. In addition, data preprocessing, model parameter tuning, and hyperparameter search further extend the training time, affecting the timeliness of forecasts. Summary of the Invention

[0006] This invention addresses the problems existing in the prior art by providing a short-term precipitation nowcasting method that integrates quantum computing and deep learning models, enabling rapid nowcasting of short-term precipitation.

[0007] To address the above technical problems, this invention provides the following technical solution: a short-term precipitation nowcasting method integrating quantum computing and deep learning models, comprising the following steps:

[0008] S1. Extract long-term historical data of precipitation and meteorological elements for the target area from the ERA5 global reanalysis data and perform preprocessing.

[0009] S2. Based on long-term historical data, construct a short-term precipitation nowcasting dataset: use the precipitation at the time of the forecast as the target quantity for prediction, and use the meteorological elements and precipitation M hours before the time of the forecast as the input features of the model.

[0010] S3. Generate a quantum circuit and use quantum computing methods to convert meteorological elements and precipitation M hours before the predicted time into high-dimensional quantum probability features of precipitation at the predicted time. Specifically, establish a quantum circuit, with the input dimension D being the total number of all input features at each time moment, and set the number of quantum circuit bits to log2(D). At the same time, set the encoding type to amplitude encoding and preset the output dimension.

[0011] By mapping input feature data to different quantum state amplitudes and embedding high-dimensional input data into the probability amplitude of quantum states, the input data is quantized.

[0012] Run the quantum circuit, measure the expected value of all qubits, and obtain the probability information of precipitation for all quantum circuits;

[0013] S4. Using the quantum circuit precipitation probability information as input and the precipitation in the next hour as output, construct and train a short-term precipitation nowcasting model. The short-term precipitation nowcasting model includes: a sequentially connected input layer, first to third convolutional layers, and a regression output layer. The first convolutional layer includes a sequentially connected first convolutional module, a first pooling layer, and a first forgetting layer; the second convolutional layer includes a sequentially connected second convolutional module, a second pooling layer, and a second forgetting layer; the third convolutional layer includes a sequentially connected third convolutional module, a global average pooling layer, and a flattening layer.

[0014] The regression output layer includes: first to third fully connected layers, third to fourth forgetting layers, first to third activation function layers, and inverse transform layer; the first fully connected layer is followed by the first activation function layer, the third forgetting layer, the second fully connected layer, the second activation function layer, the fourth forgetting layer, the third fully connected layer, the third activation function layer, and the inverse transform layer; the inverse transform layer outputs the precipitation forecast value for the next 1 hour.

[0015] Furthermore, the aforementioned step S1 includes the following sub-steps:

[0016] S1.1 Extract precipitation data and long-term historical data of meteorological elements in the target area, including: air temperature, surface air pressure, dew point temperature, wind speed, wind direction, net solar radiation at the surface, surface temperature, east-west wind speed U component, and north-south wind speed V component.

[0017] S1.2 Perform quality control and spatiotemporal alignment on the data extracted in step S1.1. For data with missing values, use linear interpolation in time or space to fill them in and perform unit conversion.

[0018] Furthermore, the aforementioned step S2 includes the following sub-steps:

[0019] S2.1 Construct a short-term precipitation nowcasting dataset. Each sample in the dataset includes an input feature vector and a target quantity. The input feature vector is defined as: X(t) = [M(t-3), M(t-2), M(t-1), P(t-3), P(t-2), P(t-1)], where M(tk) represents the meteorological element vector at time tk, k=1,2,3; P(tk) represents the precipitation at time tk, k=1,2,3; the target quantity is defined as Y(t) = P(t).

[0020] S2.2. Different normalization strategies are adopted according to the different characteristics of the target quantity and the input feature quantity. For the target quantity, minimum-maximum scaling is used to obtain the normalized precipitation Y_scaled, as shown in the following formula:

[0021] Y_scaled = (Y - Y_min) / (Y_max - Y_min) , (1)

[0022] Where Y is the precipitation at each moment, and Y_min and Y_max are the minimum and maximum precipitation values, respectively;

[0023] For the input features, Z-score normalization is used, as shown in the following formula:

[0024] X_std = (X - μ) / σ, (2)

[0025] Where X is the feature value at each time step, X_std is the normalized feature value, μ is the data mean, and σ is the standard deviation;

[0026] S2.3 Divide the short-term precipitation nowcast dataset into training set, validation set and test set according to a preset ratio.

[0027] Furthermore, the aforementioned step S3 includes the following sub-steps:

[0028] S3.1 Determine the input dimension D as the total number of all input features at each time point, i.e., D = number of meteorological feature features × m hours + number of precipitation feature features × m hours;

[0029] S3.2 Determine the number of qubits. The number of qubits n is set according to n ≥ log2(D);

[0030] S3.3. Using an amplitude coding data encoding strategy, the input feature value X_std = (X_std_1, X_std_2, ..., X_std_D) is standardized and then converted into a probability amplitude vector, ensuring that the L2 norm of the converted data is equal to 1.

[0031] S3.4. This probability amplitude vector is configured onto the ground state amplitude of the corresponding qubit through a series of quantum gate operations to obtain a complex quantum state containing all the information of the input data, and encoded in a highly entangled and nonlinear manner.

[0032] S3.5 Construct a trainable variational circuit responsible for transforming the encoded quantum state. The variational circuit consists of 4 repeating layers. The entanglement operation of each layer allows information to flow and mix between qubits, directly realizing information exchange and global correlation through quantum state.

[0033] S3.6. Measure all n qubits and return the probability distribution of all possible outcomes.

[0034] Furthermore, in the aforementioned step S3.5, each layer contains two steps: single-bit rotation and entanglement;

[0035] Specifically, single-qubit rotation involves independently and parametrically rotating all qubits to adjust their local states; entanglement is achieved by linking all qubits together using chained CNOT gates, increasing the number of states that the entire n-qubit system can represent from 2^n to 2^n. n .

[0036] Furthermore, in the aforementioned step S4, the first to third convolutional modules in the convolutional layer are used to extract local features, and the input data dimensions of the convolutional kernels are 32, 64, and 128, respectively, kernel_size=3, and padding=1 to maintain the dimensions;

[0037] The forgetting layer randomly sets the output of some neurons in the layer to 0 with a probability of dropout_rate;

[0038] The pooling layer uses kernel_size=2 for downsampling to reduce the feature dimension;

[0039] The activation function is used to ensure that the output precipitation value is not negative;

[0040] Inverse Transform Layer Performing an inverse Min-Max transform yields the final hourly precipitation forecast value in physical units. .

[0041] Furthermore, the steps for building and training a short-term precipitation nowcasting model using a CNN network, as described above, are as follows:

[0042] S4.1 Input the input quantum probability information into a CNN model with preset parameters to obtain the predicted value. ;

[0043] S4.2 Calculation using loss function The error between the actual label Y_scaled and the loss function value is used to calculate the gradient of all parameters in the CNN network using the chain rule.

[0044] S4.3 The Adam optimizer further updates all parameters based on the calculated gradients to minimize the loss function;

[0045] S4.4 Return to steps S4.1 to S4.3 and iterate through all training data for multiple cycles until the model's performance on the validation set no longer improves or reaches the preset number of training rounds.

[0046] S4.5. Evaluate the performance of the trained model using root mean square error, mean absolute error, and coefficient of determination on the test set.

[0047] Furthermore, the aforementioned use of mean squared error (MSE) as the loss function is as follows:

[0048] , (3)

[0049] Where N is the number of samples, For predicted values, This is the actual value.

[0050] Compared to existing technologies, this invention, employing the above technical solution, can achieve rapid forecasting of precipitation in the study area within the next hour on a quantum computer with lower computational burden and higher computational efficiency, mainly achieving the following effects:

[0051] 1. By utilizing quantum circuits and quantum superposition states, classical data can be mapped to a high-dimensional quantum Hilbert space, which can avoid the trap of local optima, guide optimization algorithms toward the global optimum, and improve the efficiency of model training and computation.

[0052] 2. Meteorological data is characterized by high dimensionality and multi-source heterogeneity, and its computational complexity increases exponentially with the number of features. This invention represents high-dimensional meteorological data through a linear combination of quantum states, enabling the representation and processing of high-dimensional feature spaces with exponential efficiency even with a limited number of qubits. This significantly reduces the computational burden and enables short-term precipitation forecasting under conditions of massive high-dimensional meteorological feature input.

[0053] 3. This invention maps classical data to a high-dimensional quantum Hilbert space and utilizes the superposition and entanglement properties of quantum states to extract complex high-dimensional feature correlations from input meteorological data and decode them into representations usable by subsequent classical neural networks, thereby improving the generalization ability of the model. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the quantum circuit established in this invention.

[0055] Figure 2 This is a schematic diagram of the neural network architecture of the present invention.

[0056] Figure 3 This is a comparison chart between the model of this invention and the measured values ​​on the test set.

[0057] Figure 4 This is a comparison chart of the pure classic CNN model on the test set and the measured values.

[0058] Figure 5 This is a flowchart of the technical solution of an embodiment of the present invention. Detailed Implementation

[0059] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0060] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0061] like Figure 5 As shown, this embodiment provides a short-term precipitation nowcasting method that integrates quantum computing and deep learning models, including the following steps:

[0062] S1. Extract long-term historical data of precipitation and meteorological elements for the target area from the ERA5 global reanalysis data and perform preprocessing.

[0063] In this embodiment, self-written Python code is used to download hourly data of precipitation and key meteorological elements for the study area from global ERA5 reanalysis data. ERA5 data integrates multi-source observation data from global ground observations, radiosondes, satellite remote sensing, and aircraft reports, and is coupled with numerical weather prediction model results through a four-dimensional variational assimilation system, providing high spatiotemporal resolution meteorological data globally from 1940 to the present. This invention aims to extract long-term historical hourly data of precipitation and key meteorological elements for the study area from ERA5 data with a spatial resolution of 0.25° × 0.25° (approximately 28 km × 28 km) and a temporal resolution of hourly.

[0064] The specific approach involves defining a geographic bounding box for the target region in a self-written Python code, determined by minimum / maximum longitude and latitude coordinates. For example, for a study of the monsoon region in eastern China, the bounding box could be set as lat_range = [20.0, 45.0], lon_range = [105.0, 125.0]. The extracted meteorological elements could be set as precipitation, air temperature at a height of 2 meters, surface air pressure, dew point temperature at a height of 2 meters, surface temperature, net solar radiation at the surface, wind speed and direction at a height of 10 meters, and the U-component (east-west) and V-component (north-south) wind speeds at a height of 10 meters. This self-written Python code can then automatically download and extract the corresponding ERA5 data for all grid points within the defined geographic area.

[0065] Further quality control and spatiotemporal alignment of the extracted data are performed. Missing values ​​are checked. For the very few missing points, linear interpolation in time or space can be used to fill them in. Unit conversion is also performed to ensure that all physical quantities have consistent units.

[0066] S2. Based on long-term historical data, construct a short-term precipitation nowcasting dataset: use the precipitation at the time of the proposed forecast as the target quantity for prediction, and use the meteorological elements and precipitation M hours before the time of the proposed forecast as the input features of the model.

[0067] Based on the extracted data, supervised learning samples are constructed. Each sample consists of a feature vector and a target value. The target value is defined as Y(t) = P(t), which is the precipitation at time t. The input feature value is defined as X(t) = [M(t-3), M(t-2), M(t-1), P(t-3), P(t-2), P(t-1)]. M(tk) represents the meteorological element vector at time tk, k=1,2,3; the meteorological element vector at each time includes nine meteorological element variables: air temperature, surface air pressure, dew point temperature, surface temperature, net solar radiation at the surface, wind speed, wind direction, U-wind, and V-wind. P(tk) represents the precipitation at time tk, k=1,2,3.

[0068] Different normalization strategies are adopted to address the different characteristics of the target quantity and input features. For the predicted target quantity (precipitation), min-max scaling is used to obtain the normalized precipitation Y_scaled:

[0069] Y_scaled = (Y - Y_min) / (Y_max - Y_min), (1)

[0070] Where Y is the precipitation at each time step, and Y_min and Y_max are the minimum and maximum precipitation values, respectively. This normalization method is used because precipitation is a non-negative variable that typically exhibits a long-tailed distribution (many zero values ​​and a few extreme values). Min-Max scaling can linearly map it to the [0, 1] interval, which helps stabilize the training of the quantum-classical hybrid model.

[0071] For the input features, Z-score standardization is used:

[0072] X_std = (X - μ) / σ, (2)

[0073] Where X is the feature value at each time step, X_std is the normalized feature value, μ is the data mean, and σ is the standard deviation. This normalization method is used because input meteorological features (temperature, air pressure, etc.) are usually approximately normally distributed. Z-score normalization can convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1. This method does not limit the data range, can effectively handle outliers, and can preserve the distribution information in the original data, which helps to more effectively map the data to a higher-dimensional Hilbert space in subsequent quantum-angle encoding.

[0074] After normalizing the input and output data, the entire sample set is randomly divided into training, validation, and test sets in a 70% / 15% / 15% ratio. The training set is used to train the parameters of the quantum circuit and deep learning model. The validation set is used to monitor model performance during training, adjust hyperparameters (such as learning rate and number of network layers), and perform early stopping to prevent overfitting. The test set is used to evaluate the model's final generalization ability on unseen data.

[0075] S3. Generate a quantum circuit and use quantum computing methods to convert meteorological elements and precipitation M hours before the predicted time into high-dimensional quantum probability features of precipitation at the predicted time. Specifically, establish a quantum circuit, with the input dimension D being the total number of all input features at each time moment, and set the number of quantum circuit bits to log2(D). At the same time, set the encoding type to amplitude encoding and preset the output dimension.

[0076] By mapping input feature data to different quantum state amplitudes and embedding high-dimensional input data into the probability amplitude of quantum states, the input data is quantized.

[0077] By running a quantum circuit and measuring the expected value of all qubits, the probability information of precipitation for all quantum circuits can be obtained. For example... Figure 1 As shown, the specific steps include the following:

[0078] S3.1. Establish a quantum circuit, first determining the number of qubits. In this invention, the input dimension D is the total number of all input features at each moment (9 meteorological elements × 3 hours + 1 precipitation element × 3 hours = 30 dimensions).

[0079] S3.2. The setting of the number of qubits n follows the rule n ≥ log2(D). For D=30, n≥5. Therefore, the number of qubits is determined to be 5, that is, in a 2 5 Data is processed in a 32-dimensional Hilbert space, thereby effectively constructing coupling between features through a dual quantum entanglement gate.

[0080] S3.3. An amplitude-based data encoding strategy is adopted to map the input data vector X_std = (X_std_1, X_std_2, ..., X_std_D) onto the quantum state amplitude, thereby embedding the high-dimensional input data into the probability amplitude of the quantum state and realizing the quantization of the input data. Specifically, the standardized data is first converted into a probability amplitude vector, ensuring that the L2 norm of the converted data is equal to 1.

[0081] S3.4. This probability amplitude is configured onto the ground state amplitude of the corresponding qubit through a series of quantum gate operations (such as appropriate rotation gates and controlled gates). In this way, the 30-dimensional classical eigenvector is transformed into a 5-qubit complex quantum state. This quantum state already contains all the information of the input data and is encoded in a highly entangled and nonlinear manner.

[0082] S3.5 Construct a trainable variational circuit responsible for transforming the encoded quantum state. This variational circuit consists of four repeating layers. Each layer contains two steps: Single-bit rotation: Perform independent, parameterized rotations on all qubits to adjust their local states. Entanglement: Link all qubits together using chained CNOT gates, which increases the number of states that the entire n-qubit system can represent from 2^n to 2^n. n This allows information to be processed in an exponential space. Each layer of entanglement allows information to flow and mix between qubits, directly achieving information exchange and global correlation through quantum states.

[0083] S3.6 Finally, the expected values ​​of all qubits are measured, converting the processed quantum information back to classical information. This step measures all n qubits and returns the probability distribution of all possible outcomes. Since measurement causes the quantum state to collapse to the computational ground state, the quantum circuit needs to be sampled multiple times to estimate the probability of each ground state. For the quantum circuit system in this invention, the measurement results are set to 32 possible probability distributions, which expands and reconstructs the feature space compared to the original 30-dimensional input feature space. This complete 32-dimensional probability information serves as the input data for subsequent deep learning models.

[0084] Figure 1 In the diagram, q0, q1, q2, q3, and q4 represent each qubit in the quantum circuit, which is the basic information unit in quantum computing. Each qubit can not only independently execute quantum gate operations but also interact with other qubits through entanglement, thus realizing quantum parallel computing capabilities. Rx and Rz in the diagram represent quantum gate operations rotating around the X-axis and Z-axis, respectively: X-axis rotation gate Rx(θ) = exp(-iθX / 2) and Z-axis rotation gate Rz(φ) = exp(-iφZ / 2). θ is the rotation angle parameter of the Rx gate, representing the rotation angle of the qubit around the X-axis on the Bloch sphere. φ is the rotation angle parameter of the Rz gate, representing the rotation angle of the qubit around the Z-axis on the Bloch sphere. By combining these rotation gates, precise control of quantum states can be achieved, and arbitrary single-qubit logical operations can be constructed.

[0085] S4. Using the quantum circuit's precipitation probability information as input and the precipitation for the next hour as output, construct and train a short-term precipitation nowcasting model. This model can receive the 32-dimensional probability vector output by the quantum circuit and accurately map it to the standardized precipitation for the next hour. .like Figure 2 As shown, the short-term precipitation nowcasting model includes: a sequentially connected input layer, first to third convolutional layers, and a regression output layer. The first convolutional layer includes a sequentially connected first convolutional module, a first pooling layer, and a first forgetting layer; the second convolutional layer includes a sequentially connected second convolutional module, a second pooling layer, and a second forgetting layer; and the third convolutional layer includes a sequentially connected third convolutional module, a global average pooling layer, and a flattening layer.

[0086] The regression output layer includes: first to third fully connected layers, third to fourth forgetting layers, first to third activation function layers, and inverse transform layer; the first fully connected layer is followed by the first activation function layer, the third forgetting layer, the second fully connected layer, the second activation function layer, the fourth forgetting layer, the third fully connected layer, the third activation function layer, and the inverse transform layer; the inverse transform layer outputs the precipitation forecast value for the next 1 hour.

[0087] The classic CNN part in this model consists of three convolutional layers and one regression output layer.

[0088] like Figure 2 As shown, each convolutional layer includes a convolutional module, and the structure of each convolutional module is as follows:

[0089] 1. One-dimensional convolutional layer: Uses a specified number of convolutional kernels (the input data dimensions of the first to third convolutional modules are 32, 64, and 128, respectively) to extract local features, kernel_size=3, padding=1 to preserve dimensionality;

[0090] 2. Batch Normalization Layer (BatchNorm1d): Accelerates the training process and improves model stability;

[0091] 3. ReLU Activation Function Layer: A modified linear unit. Its mathematical expression is f(x) = max(0, x). The ReLU nonlinear activation function is introduced, allowing the model to learn complex nonlinear relationships in the data. This activation function effectively alleviates the vanishing gradient problem and makes the network sparse.

[0092] In this embodiment, the first and second pooling layers use kernel_size=2 for downsampling to reduce feature dimensionality; the first and second forgetting layers are regularization techniques to prevent overfitting, with a forgetting rate d=0.2. During training, the outputs of some neurons in this layer are randomly set to 0 with a probability of dropout_rate. The forgetting rates of the third and fourth forgetting layers are set to d=0.3 and d=0.15, respectively.

[0093] The final regression output layer consists of three fully connected layers, with ReLU activation function and Dropout operation embedded in between. Finally, the third fully connected layer maps the 32-dimensional feature vector to a 1-dimensional precipitation value, which is used to output the final predicted value. The output layer also includes a Softplus activation function to ensure that the output precipitation value is not negative. Finally, for... Performing an inverse Min-Max transform yields the final hourly precipitation forecast value in physical units. Additionally, the output layer includes a Softplus activation function to ensure that the output precipitation value is not negative.

[0094] The CNN model is trained and optimized using mean squared error (MSE) as the loss function.

[0095] (3)

[0096] Where N is the number of samples. This loss function reflects the predicted value. The model calculates the magnitude of the difference between the predicted value and the true value Y_scaled_i, and imposes a higher penalty for larger prediction errors. The model is trained using the Adam adaptive moment estimator optimizer. This optimizer adaptively adjusts the learning rate for each parameter, exhibiting robustness and rapid convergence across most deep learning tasks. The training process for this CNN model is as follows: first, the input quantum probability information is fed into the CNN model with preset parameters to obtain the predicted value. Calculate using loss function The error between the model and the true label Y_scaled is calculated. Based on the loss function value, the gradients of all parameters in the CNN network are calculated using the chain rule. The Adam optimizer further updates all parameters based on the calculated gradients, minimizing the loss function. This process is repeated multiple times, iterating through all training data for multiple epochs, until the model's performance on the validation set no longer improves or reaches the preset number of training epochs.

[0097] Finally, the performance of the trained model is evaluated using root mean square error, mean absolute error, and coefficient of determination on the test set. The specific definitions of each statistical indicator are as follows:

[0098] Root mean square error: RMSE = sqrt(MSE), its unit is the same as the target variable, which can more intuitively reflect the magnitude of the prediction error.

[0099] Mean absolute error: It is less sensitive to outliers than RMSE.

[0100] Coefficient of determination: R², which represents the proportion of variance explained by the model. The closer the value is to 1, the better the model fits.

[0101] Figure 3 This paper presents the prediction results of the hourly precipitation forecast for a region surrounding a city (31.9°N, 117.3°E) from April to July 2025 using this invention, and compares them with the actual ERA5 values. The root mean square error (RMSE) is 3.63, the mean absolute error (MAE) is 1.48, and the coefficient of determination (R²) is 0.83. It can be seen that this invention can effectively capture the hourly precipitation forecast for this region at different times, and the statistical parameters also demonstrate the accuracy of the new method proposed in this invention. Figure 4 This is a comparison chart between the pure classical CNN model and the measured values ​​on the test set. The root mean square error (RMSE) is 4.38, the mean absolute error (MAE) is 1.61, and the coefficient of determination (R²) is 0.76. Compared with the results of the pure classical CNN neural network, the prediction results of the method of this invention are more accurate, and the model has better generalization ability.

[0102] Based on the above experimental results, the short-term precipitation nowcasting method provided by this invention, which integrates quantum computing and deep learning models, can quickly forecast short-term precipitation, overcome the shortcomings of existing technologies, and produce positive results, as detailed below:

[0103] 1. Quantum circuit mapping capabilities can be used to accelerate model training.

[0104] Traditional deep learning model training mainly relies on optimization algorithms such as gradient descent, which are computationally expensive and time-consuming. The quantum-classical hybrid model proposed in this invention utilizes quantum circuits to map classical data to a high-dimensional quantum Hilbert space, thereby changing the landscape of the optimization problem. This allows the model to converge faster or escape local optima under the drive of a classical optimizer (such as Adam), improving training efficiency.

[0105] 2. Can efficiently process high-dimensional feature spaces

[0106] Meteorological data often exhibits high dimensionality and multi-source heterogeneity. Traditional data-driven models are prone to the "curse of dimensionality" when processing such data, with computational complexity increasing exponentially with the number of features. Classical computers encounter computational bottlenecks when dealing with this complexity. This invention leverages the property that quantum states exist in high-dimensional Hilbert spaces, encoding classical data into specific quantum states via quantum circuits. This encoding method inherently possesses a high-dimensional advantage in representation, using only a small number of qubits to map and express the structure of high-dimensional feature spaces. This alleviates the "curse of dimensionality" problem faced by classical models, significantly reduces the computational burden, and solves the problem of short-term precipitation forecasting under conditions of massive high-dimensional meteorological data feature input.

[0107] 3. Utilize quantum superposition state mapping to obtain complex features and correlations in data, thereby improving the model's generalization ability.

[0108] This invention maps traditional data into a high-dimensional feature space where quantum superposition states exist, which can obtain more complex features and correlations in the data. This allows the model to better learn and represent complex nonlinear relationships in high-dimensional data, and can provide more informative and separable feature representations for subsequent classical neural networks, thereby improving the model's generalization ability.

[0109] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A short-term precipitation nowcast method fusing quantum computing and a deep learning model, characterized in that, The method comprises the following steps: S1, extracting long-term historical data of precipitation and meteorological elements in the target area from ERA5 global reanalysis data and preprocessing; S2, based on the long-term historical data, a short-time precipitation nowcast data set is constructed: the precipitation at the predicted time is taken as the predicted target quantity, and the meteorological elements and precipitation M hours before the predicted time are taken as the model input characteristic quantity; S3, generate a quantum circuit, and use quantum computing method to convert the meteorological elements and precipitation M hours before the predicted time into high-dimensional quantum probability characteristics of the precipitation at the predicted time; Specifically, a quantum circuit is established, the input dimension D is the total number of all input characteristics at each time, the number of quantum circuit bits is set to log2(D); At the same time, the encoding type is set to amplitude encoding, and the output dimension is preset; Map the input feature data to different quantum state amplitudes, embed the high-dimensional input data into the probability amplitude of the quantum state, and realize the quantization of the input data; Run the quantum circuit, measure the expected value of all quantum bits, and get the precipitation probability information of all quantum circuits; S4, taking the quantum circuit precipitation probability information as input and the precipitation in the future 1 hour as output, a short-time precipitation nowcast model is constructed and trained, and the short-time precipitation nowcast model comprises: an input layer, a first to third convolutional layer, and a regression output layer connected in sequence; The first convolutional layer comprises a first convolutional module, a first pooling layer and a first forget layer connected in sequence; The second convolutional layer comprises a second convolutional module, a second pooling layer and a second forget layer connected in sequence; The third convolutional layer comprises a third convolutional module, a global average pooling layer and a flattening layer connected in sequence; The regression output layer comprises: a first to third fully connected layer, a third to fourth forget layer, a first to third activation function layer, and an inverse transformation layer; The first fully connected layer is sequentially connected with the first activation function layer, the third forget layer, the second fully connected layer, the second activation function layer, the fourth forget layer, the third fully connected layer, the third activation function layer, and the inverse transformation layer; The inverse transformation layer outputs the predicted value of the precipitation in the future 1 hour; The first forget layer and the second forget layer are a regularization technique for preventing overfitting, and the forgetting rate d=0.2; During training, the output of some neurons in the layer is set to 0 with a probability of dropout_rate, and the forgetting rates of the third and fourth forget layers are set to d=0.3 and d=0.15, respectively.

2. The short-term precipitation nowcasting method of fusing quantum computing and deep learning model according to claim 1, characterized in that, Step S1 includes the following substeps: S1.1, extract long-term historical data of precipitation and meteorological elements in the target area, including: air temperature, ground pressure, dew point temperature, wind speed, wind direction, ground surface net solar radiation, ground temperature, east-west wind speed U component, north-south wind speed V component; S1.2, quality control and space-time alignment are performed on the data extracted in step S1.1, for the data with missing values, linear interpolation in time or space is used for filling, and unit conversion is performed.

3. The short-term precipitation nowcasting method of fusing quantum computing and deep learning model according to claim 1, characterized in that, Step S2 includes the following substeps: S2.1, construct a short-time precipitation nowcast dataset, each sample in the dataset includes an input feature vector and a target quantity, the input feature vector is defined as: X(t) = [M(t-3), M(t-2), M(t-1), P(t-3), P(t-2), P(t-1)], wherein M(t-k) represents the meteorological element vector at t-k time, k=1, 2, 3; P(t-k) represents the precipitation at t-k time, k=1, 2, 3; the target quantity is defined as Y(t) = P(t); S2.2, different normalization strategies are adopted according to the different characteristics of the target quantity and the input feature quantity, for the target quantity, the minimum-maximum scaling is adopted, and the normalized precipitation Y_scaled is obtained as follows: Y_scaled = (Y - Y_min) / (Y_max - Y_min), (1) wherein Y is the precipitation at each time, Y_min and Y_max are the minimum and maximum values of the precipitation respectively; For the input feature quantity, Z-score standardization is adopted as follows: X_std = (X - μ) / σ, (2) wherein X is the feature value at each time, X_std is the normalized feature value, μ is the data mean, and σ is the standard deviation; S2.3, the short-time precipitation nowcast dataset is divided into training set, validation set and test set according to the preset proportion.

4. The short-term precipitation nowcasting method of fusing quantum computing and deep learning model according to claim 1, characterized in that, Step S3 includes the following substeps: S3.1, determine the input dimension D as the total number of all input features at each time, that is, D = meteorological element feature number × m hours + precipitation element feature number × m hours; S3.2, determine the number of quantum bits, the number of quantum bits n is set as n ≥ log2(D); S3.3, adopt the amplitude coding data encoding strategy, standardize the input feature value X_std = (X_std_1, X_std_2,..., X_std_D), and then convert it into a probability amplitude vector, so that the two norm of the converted data is equal to 1; S3.4, the probability amplitude vector is configured to the ground state amplitude of the corresponding quantum bit through a series of quantum gate operations, a complex quantum state containing all the information of the input data is obtained, and it is encoded in a highly entangled and nonlinear way; S3.5, construct a trainable variational circuit responsible for transforming the encoded quantum state, which is composed of 4 repeated layers, and the entanglement operation of each layer makes the information flow and mix between quantum bits, which directly realizes information exchange and global correlation through quantum state; S3.6, measure all n quantum bits and return the probability distribution of all possible results.

5. The short-term precipitation nowcasting method fusing quantum computing and deep learning model according to claim 4, characterized in that, In step S3.5, each layer includes two steps: single-bit rotation and entanglement; Wherein, single bit rotation, specifically, independent, parameterized rotation is performed on all qubits to adjust the local state; entanglement, specifically, all qubits are associated through chain CNOT gate, so that the number of states that the entire n-qubit system can represent is increased from 2n to 2n n .

6. The short-term precipitation nowcasting method of fusing quantum computing and deep learning model according to claim 1, characterized in that, The steps of training the short-time precipitation nowcast model using the CNN network are as follows: S4.1, input the input quantum probability information into the CNN model with preset parameters to obtain a prediction value ; S4.2, calculating using loss function The error between the real label Y_scaled and the predicted label Y_pred is calculated according to the loss function value, and the gradient of all parameters in the CNN network is calculated using the chain rule. S4.3, the Adam optimizer further updates all parameters according to the calculated gradient to minimize the loss function; S4.4, return to execute step S4.1 to step S4.3, traverse all training data multiple cycles, until the performance of the model on the validation set no longer improves or reaches the preset training number of rounds; S4.5, using the root mean square error, mean absolute error and decision coefficient on the test set, evaluate the performance of the trained model.

7. The short-term precipitation nowcasting method of fusing quantum computing and deep learning model according to claim 6, characterized in that, In step S4, the first to third convolution modules in the convolution layer are used to extract local features, and the input data dimensions of the convolution kernels are 32, 64 and 128 respectively, kernel_size=3, and padding=1 to maintain the dimension; The forgetting layer randomly sets the output of some neurons in the layer to 0 with a dropout_rate probability; The pooling layer uses kernel_size=2 for down-sampling to reduce the feature dimension; The activation function is used to ensure that the output of the precipitation is not negative; Inverse transform layer pair performing an inverse Min-Max transform to obtain a final, physically unitized 1-hour precipitation forecast value .

8. The short-term precipitation nowcasting method of fusing quantum computing and deep learning model according to claim 7, characterized in that, The mean square error MSE is used as the loss function, as follows: （3）， where N is the number of samples, Y_pred is the predicted value, and Y_scaled_i is the true value.

Images