Weather forecasting method and system based on deep learning integrated model

Abstract

The invention relates to the technical field of multi-element weather forecasting, in particular to a weather forecasting method and system based on a deep learning integration model. The weather forecasting method based on the deep learning integration model comprises the following steps: acquiring weather data, and further constructing a training data set; respectively training a CNN-LSTM model, a ConvLSTM model and a CNN3D model by using the training data set; respectively processing the meteorological data input into each model by each model, and outputting three different prediction results; and respectively adopting an error-based weighted integration method and a Bayesian optimization integration method to carry out integrated prediction on the three different prediction results to obtain a final integrated prediction result. The overall scheme is simple in design and easy to implement, and the prediction precision is higher than that of a traditional scheme. According to the invention, by combining the advantages of different models, the integration method significantly improves the stability and generalization ability, and is more robust especially on complex variables such as rainfall.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-factor weather forecasting, and in particular to a weather forecasting method and system based on a deep learning integrated model. Background Art

[0002] In atmospheric science, ensemble forecasting has become a key tool for quantifying uncertainty in numerical weather prediction (NWP) and ocean modeling systems. Traditional ensemble forecasting strategies, such as the European Centre for Medium-Range Weather Forecasts (ECMWF)'s Ensemble Prediction System (EPS), primarily generate a range of possible future states of the atmosphere or ocean by perturbing initial conditions and model parameters. While this approach effectively provides probabilistic forecasts, it is computationally expensive due to the need to run multiple simulations with different inputs. For example, in high-resolution real-time forecasting scenarios, the high computational resource consumption has become a major obstacle to its practical application.

[0003] In recent years, hybrid data-driven models, which combine multi-model ensembles, have emerged as a promising approach. These models combine traditional numerical weather forecasting techniques with machine learning methods, significantly improving forecast accuracy and operational efficiency in areas such as hydrological forecasting. The following are some examples of typical model ensemble methods: 1) Using the ideal Lorenz63 and Lorenz96 models as examples, the initial perturbation set for ensemble forecasts was generated using the nonlinear local Lyapunov vector (NLLV) and growing mode propagation (BGM) methods. 2) Based on the Bayesian model averaging method, the weights of each model were determined by calculating the posterior probability, and the forecast results of multiple models were weighted averaged. 3) For wind farm wind speed and wind power forecasts, quantile regression, K-value nearest neighbor, and ensemble prediction methods were compared and analyzed. Ensemble numerical forecast members were dynamically tested and screened based on wind speed fluctuation characteristics. The selected forecast members were then subjected to quantile regression probability interval prediction. 4) Numerical weather forecasts were coupled with the Xin'anjiang Model (XAJ), the distributed hydrology-vegetation-soil model (DHSVM), and the long short-term memory (LSTM) network model. The forecast results were fused using equal weighting, unequal weighting, and BP neural network (BPNN) methods to generate ensemble forecasts. 5) Five machine learning-based deterministic precipitation prediction models are constructed: ENR, SVR, RF, XGB, and LGB. The improved stacking ensemble strategy (MSES) and the traditional Bayesian model averaging (BMA) method are used for multi-model fusion.

[0004] However, when applied to multiple meteorological factors, these methods are overly complex to design, and their accuracy, stability, and generalization capabilities, especially their robustness to complex variables, need to be improved. Therefore, those skilled in the art are in urgent need of a new weather forecasting method. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the defects existing in the above-mentioned prior art, thereby providing a weather forecasting method and system based on a deep learning integrated model.

[0006] A weather forecasting method based on a deep learning ensemble model, comprising: S1. Acquire meteorological data and further preprocess the data to construct a training dataset. Use the training dataset to train the CNN-LSTM model, ConvLSTM model, and CNN3D model. S2. Use the trained CNN-LSTM model, ConvLSTM model, and CNN3D model to process the meteorological data input to each model and output three different prediction results; S3. Use the error-based weighted integration method and the Bayesian optimization integration method to perform integrated prediction on the three different prediction results to obtain the final integrated prediction result.

[0007] Preferably, the error-based weighted integration is specifically to calculate the error of each model on a pre-constructed validation set so as to assign a greater weight to the model with stronger predictive ability.

[0008] Preferably, the expression of the ensemble prediction result generated by the error-based weighted ensemble method is: ; in, is the prediction result of the mth model at time step t, weight Calculated as: ; in, is the prediction error of the mth model at time step t, and the weight satisfy: ; Where M is the total number of models, represents the prediction error of the j-th model at time step t; It represents the integrated prediction result generated by processing the prediction results of M models using the error-based weighted integration method.

[0009] Preferably, the Bayesian optimization ensemble method determines the weights by solving the following optimization problem: ; Where L is the loss function, is the weight vector, is the optimal weight vector, is the true value at time step t, and the constraint and .

[0010] Preferably, the loss function in the Bayesian optimization ensemble method is the mean absolute error (MAE).

[0011] Preferably, meteorological data is obtained and further data preprocessing is performed, specifically including: Five meteorological elements were selected as predictor variables, and relevant data were obtained from ERA5 data in time units. The dataset was constructed and further split into training set, validation set and test set; And use minimum and maximum normalization to map the data in the dataset to values ​​between [0, 1]; The sliding window method is used to process the normalized data set to construct a time series data set that is convenient for input into various models.

[0012] A weather forecast system based on a deep learning integrated model, comprising: a data initial prediction module and an integrated prediction module connected in sequence; Data acquisition and processing module: Acquire meteorological data and further preprocess the data to construct a training dataset; use the training dataset to train the CNN-LSTM model, ConvLSTM model, and CNN3D model respectively; Data initial prediction module: Use the trained CNN-LSTM model, ConvLSTM model, and CNN3D model to process the meteorological data input to each model and output three different prediction results; Integrated prediction module: The error-based weighted integration method and the Bayesian optimization integration method are used to perform integrated prediction on three different prediction results to obtain the final integrated prediction result.

[0013] The technical solution of the present invention has the following advantages: The overall solution of the present invention is simple in design, easy to implement, and has higher prediction accuracy than traditional solutions. By combining the advantages of different models, the present invention significantly improves the stability and generalization ability of the integration method, especially in complex variables such as precipitation. Whether it is an integration method based on error weighting or Bayesian optimization, the full variable / single variable of the integrated model is better than the single deep learning model in most indicators. In the Bayesian optimization integration method, the overall prediction effect of the single variable integration performs best in all three indicators, especially in precipitation. The integration effect of the present invention that independently optimizes the weight of each variable is better, and the overall MAE is reduced by 17.26% and 7.51% respectively in the two integration methods compared with the full variable integration, especially in precipitation, the error reduction is more than 60% and 30% respectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0015] Figure 1 This is a schematic diagram of the implementation process of the first embodiment of the present invention; Figure 2 This is the structure diagram of the CNN-LSTM model; Figure 3 This is the relationship between the real 10-meter zonal wind component and time; Figure 4 This is the relationship diagram between the 10-meter zonal wind volume and time prediction by the CNN-LSTM model; Figure 5 Schematic diagram of CNN-LSTM model prediction error; Figure 6 Bayesian optimization ensemble method integrates 10-meter zonal wind component prediction map under single variable; Figure 7 Schematic diagram of the ensemble prediction error of the Bayesian optimization ensemble method under single variable. DETAILED DESCRIPTION

[0016] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0017] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0018] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0019] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0020] Example 1 First, it should be noted that the method of this embodiment proposes a new integration approach. While integrating multiple models is a trend in traditional methods, integrating CNN-LSTM, ConvLSTM, and CNN3D models is not yet common in the field of weather forecasting. This is because researchers generally believe that in practical applications, the appropriate model for integration is typically selected based on the characteristics of the specific task to achieve optimal performance. For example, in weather data forecasting, multiple CNN-LSTM or ConvLSTM models may be integrated to improve prediction accuracy; in video analysis, a CNN3D model may be integrated with other specially designed video models. These existing integration solutions can already meet the needs of practical applications to a certain extent, so there is no need to integrate CNN-LSTM, ConvLSTM, and CNN3D models. However, while the solutions in the prior art can meet the needs of practical applications to a certain extent, the accuracy, robustness, and simplicity of the overall solution still need to be improved.

[0021] Therefore, this embodiment discloses a weather forecasting method based on a deep learning integrated model, including: S1. Acquire meteorological data and further preprocess the data to construct a training dataset. Use the training dataset to train the CNN-LSTM model, ConvLSTM model, and CNN3D model. S2. Use the trained CNN-LSTM model, ConvLSTM model, and CNN3D model to process the meteorological data input to each model and output three different prediction results; S3. Use the error-based weighted integration method and the Bayesian optimization integration method to perform integrated prediction on the three different prediction results to obtain the final integrated prediction result.

[0022] Specifically: like Figure 1This is a schematic diagram of the implementation process of this embodiment, which involves data preprocessing, deep learning prediction, integrated prediction and prediction effect evaluation; and in the integrated prediction, single variable integration and multivariate weighted integration are also considered.

[0023] Acquire meteorological data and perform further data preprocessing, including: Five meteorological elements were selected as predictor variables, and relevant data were obtained from the ERA5 data on a time-based basis. A dataset was constructed and further split into a training set, a validation set, and a test set. Specifically, the data used in this example were from the ERA5 reanalysis data with a resolution of 0.25°. The time period was from 00:00 on July 1, 2015, to 23:00 on August 15, 2015, with a sampling interval of 1 hour. Grid data was selected from the three northeastern provinces between 40° and 53° north latitude and 120° and 135° east longitude. Five meteorological elements from the ERA5 data were selected as predictor variables: the 10-meter zonal wind component, the 10-meter meridional wind component, the 2-meter temperature, the 2-meter dew point temperature, and the total precipitation. The first 80% of the July meteorological data was used as the training set, 20% as the validation set, and the meteorological data from August 1 to 15 as the test set.

[0024] In order to eliminate the influence of differences in dimension and magnitude and make different variables comparable, the minimum and maximum normalization is used to map the data in the dataset to values ​​between [0, 1]. The calculation expression is: ; in represents the normalized value, Represents the value of a variable, and Represent variables respectively; The sliding window method is used to process the normalized data set to construct a time series data set that is convenient for input to each model. The window size used is 12 time steps, and the input feature matrix ; Prediction target , ; is the multivariate observation value at time t, each observation value contains all the selected meteorological elements.

[0025] in, .

[0026] The trained CNN-LSTM model, ConvLSTM model, and CNN3D model are used to process the meteorological data input into each model, and three different prediction results are output. Specifically: The CNN-LSTM model first reshapes multivariate data into a form suitable for CNN processing, extracts spatial features through two CNN layers, and then the reorganization layer inputs the reorganized time series information into the LSTM layer to process time dependencies, and cooperates with the Dropout layer to prevent overfitting. Finally, the fully connected layer generates the prediction results. This CNN-LSTM model is more suitable for multivariate weather forecasts with spatiotemporal dependencies. The CNN-LSTM model architecture is shown in the figure below. Figure 2 As shown in the figure, the prediction results of the variable 10-meter zonal wind component are as follows Figure 4-6 shown.

[0027] The ConvLSTM model processes each variable using a ConvLSTM2D layer, which captures both spatial and temporal dependencies. The ConvLSTM model first reshapes the input data, then extracts spatiotemporal features using the ConvLSTM2D layer. BatchNormalization and Activation layers are used to enhance model training, and finally, a 1x1 convolution is used to generate predictions. This ConvLSTM model is particularly well-suited for processing spatially continuous meteorological data sequences.

[0028] The CNN3D model uses a three-dimensional convolutional neural network to process multivariate spatiotemporal data. It first flattens all variables and time steps. It then uses a 3D convolutional layer to simultaneously process features in both the temporal and spatial dimensions. After average pooling across both the variables and the temporal dimension, it uses a 2D convolutional layer to further process spatial features. Finally, a 1x1 convolution layer generates predictions. This CNN3D model effectively captures the spatiotemporal correlations in the data. A comparative analysis of the three deep learning prediction models is shown in the following table: Table 1 Comparison of deep learning models

[0029] Among them, the error-based weighted integration is specifically to calculate the error of each model on the pre-built validation set so that the model with stronger predictive ability is assigned a larger weight.

[0030] The expression of the ensemble prediction result generated by the error-based weighted ensemble method is: ; in, is the prediction result of the mth model at time step t, weight Calculated as: ; in, is the prediction error of the mth model at time step t, and the weight satisfy: ; Where M is the total number of models, represents the prediction error of the j-th model at time step t; It represents the integrated prediction result generated by processing the prediction results of M models using the error-based weighted integration method.

[0031] The average weights of the models and variables obtained by the weighted ensemble prediction based on the error are shown in the following table: Table 2 Average weights of model-variable correspondences: error-based weighted integration method

[0032] The Bayesian optimization ensemble method determines the weights by solving the following optimization problem: ; Where L is the loss function, is the weight vector, is the optimal weight vector, is the true value at time step t, and the constraint and The loss function in the Bayesian optimization ensemble method is the mean absolute error (MAE). It should be noted that the Bayesian optimization ensemble uses the Bayesian optimization algorithm to find the optimal model weight combination. By defining the objective function to minimize the MAE, it searches in the weight space to find the globally optimal weight configuration.

[0033] The prediction results of the 10-meter zonal wind component using single variable integration are as follows: Figure 4 、 6 -7. The average weights of the models and variables obtained by Bayesian optimization ensemble prediction are shown in the following table: Table 3. Average weights of model-variable pairs: Bayesian optimization ensemble method

[0034] In addition, this embodiment also includes prediction effect evaluation; The comparative analysis evaluated the model performance based on two ensemble methods and full / single variable ensemble methods. Full variable ensemble processes all variables simultaneously, generating a set of weights for each time step that are applied to all variables simultaneously. This method considers any interrelationships between variables. Single variable ensemble processes each variable individually, generating a set of weights specific to that variable for each time step. This method ignores interrelationships between variables and focuses solely on the prediction quality of the current variable.

[0035] Mean absolute error MAE, root mean square error RMSE and Pearson correlation coefficient r were used as prediction effect evaluation indicators; The MSE, RMSE, and correlation coefficients of the model-variables based on different integration methods are as follows: Table 4. Evaluation results of model-variable correspondences: error-based weighted ensemble method

[0036] Table 5. Evaluation results of model-variable correspondence: Bayesian optimization ensemble method

[0037] As can be seen from Tables 4 and 5 above, by combining the strengths of different models, ensemble methods significantly improve stability and generalization, particularly for complex variables such as precipitation. Whether based on error weighting or Bayesian optimization, the ensemble model (all variables / single variable) outperforms single deep learning models on most metrics. Among the Bayesian optimization ensemble methods, the single-variable ensemble achieved the best overall prediction performance across all three metrics, with a particularly strong performance on precipitation. Ensembles that independently optimize the weights of each variable achieved even better results, with overall MAE reductions of 17.26% and 7.51% for both ensemble methods compared to the full-variable ensemble. In particular, the error reduction for precipitation exceeded 60% and 30%, respectively.

[0038] Example 2 This embodiment discloses a weather forecast system based on a deep learning integrated model, comprising: a data initial prediction module and an integrated prediction module connected in sequence; Data acquisition and processing module: Acquire meteorological data and further preprocess the data to construct a training dataset; use the training dataset to train the CNN-LSTM model, ConvLSTM model, and CNN3D model respectively; Data initial prediction module: Use the trained CNN-LSTM model, ConvLSTM model, and CNN3D model to process the meteorological data input to each model and output three different prediction results; Integrated prediction module: The error-based weighted integration method and the Bayesian optimization integration method are used to perform integrated prediction on three different prediction results to obtain the final integrated prediction result.

[0039] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will readily appreciate that other variations or modifications based on the above descriptions are possible. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A weather forecasting method based on a deep learning integrated model, characterized in that: include: S1. Acquire meteorological data and further preprocess the data to construct a training dataset; Use the training dataset to train the CNN-LSTM model, ConvLSTM model, and CNN3D model respectively; S2. Use the trained CNN-LSTM model, ConvLSTM model, and CNN3D model to process the meteorological data input to each model and output three different prediction results; S3. Use the error-based weighted integration method and the Bayesian optimization integration method to perform integrated prediction on the three different prediction results to obtain the final integrated prediction result.

2. A weather forecast method based on a deep learning integrated model according to claim 1, characterized in that: Error-based weighted ensemble specifically calculates the error of each model on a pre-constructed validation set so that models with stronger predictive power are assigned greater weights.

3. A weather forecasting method based on a deep learning integrated model according to claim 1, characterized in that: The expression of the ensemble prediction result generated by the error-based weighted ensemble method is: ； in, is the prediction result of the mth model at time step t, weight Calculated as: ； in, is the prediction error of the mth model at time step t, and the weight satisfy: ； Where M is the total number of models, represents the prediction error of the j-th model at time step t; It represents the integrated prediction result generated by processing the prediction results of M models using the error-based weighted integration method.

4. A weather forecasting method based on a deep learning integrated model according to claim 1, characterized in that: The Bayesian optimization ensemble method determines the weights by solving the following optimization problem: ； Where L is the loss function, is the weight vector, is the optimal weight vector, is the true value at time step t, and the constraint and .

5. A weather forecasting method based on a deep learning integrated model according to claim 4, characterized in that: The loss function in the Bayesian optimization ensemble method is the mean absolute error (MAE).

6. A weather forecast method based on a deep learning integrated model according to claim 1, characterized in that: Acquire meteorological data and perform further data preprocessing, including: Five meteorological elements were selected as predictor variables, and relevant data were obtained from ERA5 data in time units. The dataset was constructed and further split into training set, validation set and test set; And use minimum and maximum normalization to map the data in the dataset to values ​​between [0, 1]; The sliding window method is used to process the normalized data set to construct a time series data set that is convenient for input into various models.

7. A weather forecast system based on a deep learning integrated model, characterized by comprising: A data initial prediction module and an integrated prediction module connected in sequence; Data acquisition and processing module: obtains meteorological data and further preprocesses the data to construct a training data set; Use the training dataset to train the CNN-LSTM model, ConvLSTM model, and CNN3D model respectively; Data initial prediction module: Use the trained CNN-LSTM model, ConvLSTM model, and CNN3D model to process the meteorological data input to each model and output three different prediction results; Integrated prediction module: The error-based weighted integration method and the Bayesian optimization integration method are used to perform integrated prediction on three different prediction results to obtain the final integrated prediction result.

Images