A short-medium-long term precipitation forecast error intelligent correction method

Abstract

The application discloses a short-medium-long-term precipitation forecast error intelligent correction method, comprising the following steps: obtaining historical precipitation forecast data and measured meteorological data of a forecast area, and performing pretreatment and error characteristic analysis to construct a training data set; constructing and optimizing a time-effect adaptive precipitation forecast correction model, including a BO-LSTM correction model and a BO-RF correction model; determining a double-model switching critical forecast period between the LSTM model and the RF correction model, and then constructing a time-effect adaptive switching mechanism; obtaining original precipitation forecast data to be corrected of the forecast area, identifying a forecast period, and automatically matching a corresponding correction model through the time-effect adaptive switching mechanism to perform hierarchical correction on short, medium and long-term precipitation forecast data, and obtain corrected precipitation forecast values. The application realizes model adaptive switching based on error quantization rules, effectively makes up for the prediction defects of a single model, and can be widely applied to regional 1-40-day short, medium and long-term precipitation forecast correction and drought and flood disaster warning scenes.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, specifically relating to an intelligent correction method for short-, medium-, and long-term precipitation forecasts. Background Technology

[0002] Precipitation forecasting is a crucial support for drought and flood early warning, irrigation scheduling, and food security. As forecast failures lengthen, the systematic bias of numerical weather prediction products increases significantly: short-term (1-3 days) forecasts have relatively high accuracy; medium-term (4-15 days) forecasts begin to decline in accuracy, exhibiting systematic underestimation; and long-term (16-40 days) errors accumulate, leading to a substantial reduction in predictability.

[0003] Existing error correction methods are mainly divided into two categories: one is traditional statistical methods (such as decreasing average and stepwise regression), which are computationally simple but lack the ability to characterize nonlinear errors; the other is machine learning methods (such as Long Short-Term Memory (LSTM) and Random Forest (RF), which typically employ standard model structures. However, in existing technologies, a single model struggles to maintain optimal performance across the entire time frame: LSTM excels at capturing short-term time-series dependencies, but its long-term predictions are susceptible to error accumulation; RF is robust to high-dimensional features and does not accumulate errors, but its short-term predictions may be inferior to time-series models. Currently, there is a lack of a method that can quantitatively utilize the variation of error over the forecast period, automatically switch models at appropriate critical points, and smoothly fuse them. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides an intelligent correction method for short-, medium-, and long-term precipitation forecast errors. This method relies on error quantification laws to achieve adaptive model switching, effectively compensating for the forecast deficiencies of a single model. It can be widely applied to regional 1-40 day short-, medium-, and long-term precipitation forecast correction and drought and flood disaster early warning scenarios.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: a smart correction method for short-, medium-, and long-term precipitation forecast errors, comprising the following steps: S100. Obtain historical precipitation forecast data and measured meteorological data for the forecast area, perform preprocessing and error feature analysis, and construct a training dataset; S200. Based on the training dataset, construct and optimize the time-adaptive precipitation forecast correction model, including the BO-LSTM correction model for short-term precipitation error correction and the BO-RF correction model for medium- to long-term precipitation error correction. S300, determine the critical forecast period for switching between the LSTM model and the RF correction model, and then construct a timely adaptive switching mechanism; S400: Obtain the original precipitation forecast data to be corrected for the forecast area, identify the forecast lead time, and automatically match the corresponding correction model through the time-adaptive switching mechanism to perform stratified correction on the short, medium and long-term precipitation forecast data to obtain the corrected precipitation forecast value.

[0006] Further, step S100 includes the following sub-steps: S101. Obtain historical weather forecast products and corresponding measured meteorological data from multiple meteorological stations within the target forecast area; The historical weather forecast products include forecasted rainfall, forecasted maximum temperature, forecasted minimum temperature, and forecasted sunshine hours, while the measured meteorological data is the measured daily rainfall. S102. Clean the collected data, remove missing data, and perform unified spatiotemporal resolution matching to eliminate abnormal noise data. Convert the forecast sunshine hours into quantitative values ​​and construct a standardized historical rainfall forecast sample dataset. S103. Based on the standardized historical rainfall forecast sample dataset, perform error feature analysis and construct a training dataset; The error characteristic analysis includes exploring the spatiotemporal distribution patterns of precipitation forecast errors, statistically analyzing the underreporting and false alarm rates of forecast models under different precipitation intensity levels, and comparing the differences in precipitation forecast errors among different sub-regions within the forecast area.

[0007] Further, step S200 includes the following sub-steps: S201. Perform Z-score standardization on the meteorological data in the training dataset; The meteorological data includes forecasted rainfall, forecasted maximum temperature, forecasted minimum temperature, forecasted sunshine hours, and measured rainfall. S202. Based on the standardized training dataset, construct the LSTM base model and the RF base model respectively; S203. Using a Bayesian optimization framework, with the minimization of the root mean square error of K-fold cross-validation as the objective function, a global adaptive search is performed on the hyperparameters of the LSTM base model and the RF base model, respectively. S204. During the iterative search process, the optimal hyperparameter combination is obtained based on the minimum cross-validation error of the Bayesian optimization output, thereby obtaining the BO-LSTM correction model for short-term precipitation error correction and the BO-RF correction model for medium- to long-term precipitation error correction.

[0008] Furthermore, in S202, for the LSTM base model, the predicted rainfall, predicted minimum temperature and predicted sunshine hours in the standardized training dataset are used as input features, and the measured rainfall is used as the prediction target. For the RF basic model, the multi-dimensional meteorological data in the standardized training dataset is used as input features, and the measured rainfall is used as the prediction target; wherein, the multi-dimensional meteorological data includes the forecast rainfall, the forecast maximum temperature, the forecast maximum temperature and the forecast sunshine duration.

[0009] Furthermore, in step S202, during the construction of the LSTM base model, the Adam optimizer is used, initial hyperparameters are configured for iterative training, and an early stopping mechanism is introduced to prevent model overfitting; the hyperparameters in the LSTM base model include learning rate, number of hidden units, Dropout rate, and backtracking window. In the process of constructing the RF basic model, multiple decision trees are integrated by bootstrapping and initial hyperparameters are configured. The model structure is initially adjusted by combining out-of-bag error. The hyperparameters in the RF basic model include the number of decision trees, the maximum number of splits, the minimum number of leaf nodes, and the feature material strategy.

[0010] Furthermore, in S203, for the LSTM base model and the RF base model, in the global adaptive search process using the Bayesian optimization framework, the search space is set as a continuous or discrete interval as a hyperparameter based on prior knowledge, and the posterior distribution is iteratively updated through a Gaussian process surrogate model.

[0011] Furthermore, in S204, the method for determining the BO-LSTM correction model is as follows: After obtaining the optimal hyperparameter combination, the data in the training dataset are divided into a training subset and a validation subset in chronological order. The LSTM base model is retrained on the training subset using the optimal hyperparameter combination, and overfitting is monitored using the validation subset. At the same time, L2 regularization and early stopping mechanism are applied to obtain the BO-LSTM correction model for short-term precipitation error correction. The method for determining the BO-RF correction model is as follows: The RF base model is retrained using the optimal hyperparameter combination and concentrated data from the training dataset, and the model stability is improved by combining the Bagging ensemble strategy, thus obtaining the BO-RF correction model for medium- to long-term precipitation error correction.

[0012] Furthermore, step S300 includes the following sub-steps: S301. Use the BO-LSTM correction model and the BO-RF correction model respectively to model and test the unified training dataset, and calculate the corresponding error indexes for all short, medium and long-term forecast periods in batches. S302. For the error indices of the BO-LSTM correction model and the BO-RF model under the same forecast period, the forecast duration corresponding to the first time when the error index of the BO-LSTM correction model is greater than the error index of the BO-RF correction model is taken as the critical forecast period for switching between the two models. S303. Based on the critical forecast period for dual-model switching, a time-adaptive switching mechanism is constructed, including: When the forecast period is less than the critical forecast period for switching between the two models, it is determined to be a short forecast period, and the BO-LSTM correction model is called to correct the precipitation forecast error. When the forecast period is longer than the critical forecast period for switching between the two models, it is determined to be a medium- to long-term forecast, and the BO-RF correction model is called to correct the precipitation forecast error.

[0013] Furthermore, the error index is the precipitation forecast RMSE.

[0014] The beneficial effects of this invention are as follows: (1) Scientific decision-making based on the quantitative laws of error: This invention analyzes the error characteristics of historical precipitation forecast data, revealing a significant cumulative effect of forecast errors as the lead time increases. Based on this, models are constructed for different lead times. Furthermore, by systematically analyzing the performance differences between BO-LSTM and BO-RF under different lead times, the switching critical point is quantitatively determined, avoiding empirical settings and improving the reliability and reproducibility of the correction strategy.

[0015] (2) Improved accuracy across all time periods: In this invention, the sensitivity of BO-LSTM to time-series dependencies is utilized in the short term (1-3 days), while the robustness of BO-RF to high-dimensional features and the non-accumulation of errors are utilized in the medium to long term (4-40 days). Based on the error statistical characteristics of different time periods, an adaptive segmented correction framework for prediction-dependent errors is proposed. Attached Figure Description

[0016] Figure 1 The flowchart of the intelligent correction method for short-, medium-, and long-term precipitation forecast errors provided by this invention is shown.

[0017] Figure 2 This is a schematic diagram showing the results of rainfall forecast correction before and after the correction provided by the BO-LSTM correction model of this invention.

[0018] Figure 3 This is a schematic diagram showing the results of rainfall forecast correction before and after correction based on the BO-RF correction model provided by the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0020] This invention provides an intelligent correction method for short-, medium- and long-term precipitation forecast errors. Based on quantitative analysis of the changes in forecast error with the lead time, the optimal critical lead time for switching between the two models is determined. This enables adaptive correction using LSTM for short-term (1-3 days) and RF for medium- and long-term (4-40 days), and systematically optimizes the model hyperparameters, thereby significantly improving the accuracy of precipitation forecasts across the entire time range.

[0021] refer to Figure 1 A smart correction method for short-, medium-, and long-term precipitation forecast errors includes the following steps: S100. Obtain historical precipitation forecast data and measured meteorological data for the forecast area, perform preprocessing and error feature analysis, and construct a training dataset; S200. Based on the training dataset, construct and optimize the time-adaptive precipitation forecast correction model, including the BO-LSTM correction model for short-term precipitation error correction and the BO-RF correction model for medium- to long-term precipitation error correction. S300, determine the critical forecast period for switching between the LSTM model and the RF correction model, and then construct a timely adaptive switching mechanism; S400: Obtain the original precipitation forecast data to be corrected for the forecast area, identify the forecast lead time, and automatically match the corresponding correction model through the time-adaptive switching mechanism to perform stratified correction on the short, medium and long-term precipitation forecast data to obtain the corrected precipitation forecast value.

[0022] S100 of this embodiment of the invention includes the following sub-steps: S101. Obtain historical weather forecast products and corresponding measured meteorological data from multiple meteorological stations within the target forecast area; The historical weather forecast products include forecasted rainfall, forecasted maximum temperature, forecasted minimum temperature, and forecasted sunshine duration, while the measured meteorological data is the measured daily rainfall; specifically, the time span of the above collected data is no less than 5 years; S102. Clean the collected data, remove missing data, and perform unified spatiotemporal resolution matching to eliminate abnormal noise data. Convert the forecast sunshine hours into quantitative values ​​and construct a standardized historical rainfall forecast sample dataset. S103. Based on the standardized historical rainfall forecast sample dataset, perform error feature analysis and construct a training dataset; The error feature analysis includes uncovering the spatiotemporal distribution patterns of precipitation forecast errors, statistically analyzing the false negative and false alarm rates of forecast models under different precipitation intensity levels, and comparing the differences in precipitation forecast errors among different sub-regions within the forecast area (such as the north and south regions of the Yellow River main stream). In this embodiment, through error feature analysis, it is clarified that the forecast error exhibits a significant cumulative effect with the lead time, with relatively small short-term forecast errors and gradually increasing medium- and long-term forecast errors. Based on the above analysis, the training dataset is constructed according to the lead time dimension for subsequent training and validation of BO-LSTM and BO-RF models, while also providing data support for model selection and time-adaptive switching mechanisms, thereby achieving hierarchical correction of short-term and medium- to long-term precipitation forecasts. Specifically, through the above error feature analysis, full-dimensional error feature extraction is completed, providing data support for subsequent model training and error correction.

[0023] In S200 of this embodiment of the invention, an LSTM model and a RF random forest base model are built. The hyperparameters of the two models are optimized by Bayesian optimization (BO), and the models are trained separately. At the same time, a unified preprocessing representation is used for the input data to improve the model correction accuracy and generalization ability.

[0024] Specifically, S200 in this embodiment of the invention includes the following sub-steps: S201. Perform Z-score standardization on the meteorological data in the training dataset; The meteorological data includes forecasted rainfall, forecasted maximum temperature, forecasted minimum temperature, forecasted sunshine hours, and measured rainfall; S202. Based on the standardized training dataset, construct the LSTM base model and the RF base model respectively; S203. Using a Bayesian optimization framework, with the minimization of the root mean square error of K-fold cross-validation as the objective function, a global adaptive search is performed on the hyperparameters of the LSTM base model and the RF base model, respectively. S204. During the iterative search process, the optimal hyperparameter combination is obtained based on the minimum cross-validation error of the Bayesian optimization output, thereby obtaining the BO-LSTM correction model for short-term precipitation error correction and the BO-RF correction model for medium- to long-term precipitation error correction.

[0025] In S201 of this embodiment, the process of performing Z-score standardization on meteorological feature data is as follows: Select the maximum value X of each meteorological feature data in the training dataset. max With minimum value X minThe standardized data were calculated. : Calculate the standardized data in the training dataset. Standard deviation With variance The training dataset is standardized and normalized to obtain a standardized training dataset with a mean of 0 and a variance of 1, which is used for subsequent training of the LSTM and RF base models; among which, .

[0026] Furthermore, for the standardized training dataset, k-fold cross-validation is used to divide it into multiple training-validation subset pairs. Each set is used as the validation set in turn, and the remaining subsets are used for training, so as to make full use of the limited samples and improve the stability of hyperparameter evaluation.

[0027] In S202 of this embodiment, for the LSTM base model, the predicted rainfall, predicted minimum temperature and predicted sunshine hours in the standardized training dataset are used as input features, and the measured rainfall is used as the prediction target. For the basic RF model, the multi-dimensional meteorological data in the standardized training dataset is used as input features, and the measured rainfall is used as the prediction target. The multi-dimensional meteorological data includes forecasted rainfall, forecasted maximum temperature, forecasted minimum temperature, and forecasted sunshine hours.

[0028] In S202 of this embodiment, during the construction of the LSTM base model, the Adam optimizer is used, initial hyperparameters are configured for iterative training, and an early stopping mechanism is introduced to prevent model overfitting; wherein, the hyperparameters in the LSTM base model include learning rate, number of hidden units, Dropout rate, and backtracking window; In the process of building the basic RF model, multiple decision trees are integrated through bootstrapping sampling and initial hyperparameters are configured. The model structure is initially adjusted by combining out-of-bag error. The hyperparameters in the basic RF model include the number of decision trees, the maximum number of splits, the minimum number of leaf nodes, and the feature material strategy.

[0029] In S203 of this embodiment, for the LSTM base model and the RF base model, in the global adaptive search process using the Bayesian optimization framework, the search space is set as a continuous or discrete interval as a hyperparameter based on prior knowledge, and the posterior distribution is iteratively updated through a Gaussian process surrogate model.

[0030] Specifically, for the LSTM base model, the maximum number of evaluations, the initial number of random sampling points, the type of acquisition function, and the exploration-utilization balance parameter are set, with the optimization objective being to minimize the root mean square error of k-fold cross-validation.

[0031] In S204 of this embodiment, the method for determining the BO-LSTM correction model is as follows: After obtaining the optimal hyperparameter combination, the data in the training dataset are divided into a training subset and a validation subset in chronological order. The LSTM base model is retrained on the training subset using the optimal hyperparameter combination, and overfitting is monitored using the validation subset. At the same time, L2 regularization and early stopping mechanism are applied to obtain the BO-LSTM correction model for short-term precipitation error correction. The method for determining the BO-RF correction model is as follows: The RF base model is retrained using the optimal hyperparameter combination and concentrated data from the training dataset, and the model stability is improved by combining the Bagging ensemble strategy, thus obtaining the BO-RF correction model for medium- to long-term precipitation error correction.

[0032] S300 of this embodiment of the invention includes the following sub-steps: S301. Use the BO-LSTM correction model and the BO-RF correction model respectively to model and test the unified training dataset, and calculate the corresponding error indexes for all short, medium and long-term forecast periods in batches. In a preferred embodiment, the error index is the precipitation forecast RMSE; S302. For the error indices of the BO-LSTM correction model and the BO-RF model under the same forecast period, the forecast duration corresponding to the first time when the error index of the BO-LSTM correction model is greater than the error index of the BO-RF correction model is taken as the critical forecast period for switching between the two models. S303. Based on the critical forecast period for dual-model switching, a time-adaptive switching mechanism is constructed, including: When the forecast period is less than the critical forecast period for switching between the two models, it is determined to be a short forecast period, and the BO-LSTM correction model is called to correct the precipitation forecast error. When the forecast period is longer than the critical forecast period for switching between the two models, it is determined to be a medium- to long-term forecast, and the BO-RF correction model is called to correct the precipitation forecast error.

[0033] In one specific embodiment, after extensive experimental verification, the critical prediction period Tc for dual-model switching in this invention is set to 3 days.

[0034] In S400 of this embodiment of the invention, the original numerical precipitation forecast data to be corrected for the target area is obtained, the forecast lead time is identified, and the optimal correction model is automatically matched through the above-mentioned time-adaptive switching mechanism. The short, medium and long-term precipitation forecast data for 1 to 40 days are corrected in layers, and finally, a high-precision, adaptive corrected precipitation forecast value is output.

[0035] In one specific embodiment of the present invention, taking the Shandong Yellow River Irrigation Area as an example, the precipitation forecast data is corrected.

[0036] This embodiment is based on daily forecast data and corresponding measured daily rainfall from 44 meteorological stations within the irrigation area from January 1, 2019 to December 31, 2023. The forecast data includes forecast rainfall, forecast minimum temperature, and forecast sunshine hours, while the measured data is the measured daily rainfall at the corresponding time. Data from 2019 to 2022 (a total of four years) is used as the training set, and the entire year of 2023 is used as the test set. The specific implementation steps are as follows.

[0037] Step 1: Extraction of Spatiotemporal Features of Historical Precipitation Forecast Errors and Construction of Dataset First, the collected historical data is preprocessed by removing missing values, smoothing outomas, standardizing the time resolution of all data to a daily scale, and converting the forecast sunshine duration from qualitative to quantitative values. Then, Z-score standardization is applied to all input features.

[0038] Based on preprocessed data, the spatiotemporal distribution of forecast errors under different lead times of 1 to 40 days was analyzed. The calculation results show that as the lead time increases from 1 day to 40 days, the overall root mean square error (RMSE) gradually increases from approximately 8.33 mm to approximately 9.82 mm. Simultaneously, the missed prediction rate and false alarm rate were statistically analyzed under different precipitation intensity levels (no rain, light rain, moderate rain, heavy rain and above), revealing that forecast performance varies significantly with rainfall intensity. The missed prediction rate is highest under no rain conditions (0.63), while the false alarm rate exceeds 0.83 under moderate rain and above conditions. Comparing the error differences between the two sub-regions north and south of the Yellow River main stream within the irrigation area, it was found that the forecast accuracy is slightly higher in the area north of the Yellow River main stream than in the area south of it. The above error characteristic analysis provides data support for subsequent model training and correction strategies.

[0039] Step 2: Construction of the BO-LSTM correction model based on Bayesian optimization According to the method in S200 of this invention, a BO-LSTM model suitable for short-forecast precipitation error correction is constructed.

[0040] First, the dataset is preprocessed: missing values ​​and outlier noise points are removed, the time resolution is standardized to a daily scale, and the predicted sunshine hours are converted into quantitative values. Next, Z-score standardization is used to process the input features.

[0041] In terms of sample construction, the input features for each time step are the forecasted rainfall, forecasted minimum temperature, and forecasted sunshine hours for all 44 stations, totaling 132 features. The 132-dimensional feature vectors from the previous L consecutive days are used as the input sequence (sequence length L), and the measured rainfall (44 dimensions) for each station on day L+1 (i.e., the next day) is used as the output target to construct the training sample set. The backtracking window L is the hyperparameter to be optimized, with a value ranging from 20 to 60 days (integer).

[0042] To automatically find the optimal hyperparameters, this invention introduces a Bayesian optimization framework. The prior search space for hyperparameters is set as follows: backtracking window 20–60 (integers), number of LSTM hidden units 64–128 (integers), dropout rate 0.3–0.5 (real numbers), initial learning rate 0.001–0.05 (using logarithmic transformation, real numbers), mini-batch size 32–128 (integers), and number of training epochs 50–150 (integers). Bayesian optimization uses the root mean square error (RMSE) of 5-fold cross-validation as the objective function, with the following parameter configurations: maximum number of objective function evaluations 50, initial number of random sampling points set to default, expected improvement as the sampling function, exploration ratio set to 0.5, and Gaussian process as the surrogate model. In each objective function evaluation, the training set (2019–2022) is divided into 5 folds in chronological order, with one fold used as the validation set in turn, and the remaining 4 folds used to train the LSTM model. The LSTM network structure consists of a sequence input layer, an LSTM layer (output mode is last), a Dropout layer, a fully connected layer, and a regression output layer. Training uses the Adam optimizer with mean squared error (MSE) as the loss function, incorporating L2 regularization (coefficient 0.01) and early stopping (validation set patience is 5). For each set of hyperparameters, the RMSE on each fold validation set is calculated and averaged as the objective function value for that set of hyperparameters. The posterior distribution is updated iteratively through a Gaussian process, and the algorithm automatically converges to the optimal hyperparameter combination that minimizes the cross-validation RMSE after 50 evaluations.

[0043] After obtaining the optimal hyperparameters, the original training data (2019-2022) was divided chronologically into a 90% training set and a 10% validation set (the validation set was taken from the last 10% of the time series to ensure the chronological order was not disrupted). The LSTM model was retrained on the training set using the optimal hyperparameters. Overfitting was monitored using the validation set during training, and an early stopping mechanism (patience value of 5) and L2 regularization (coefficient 0.01) were employed. After early stopping was triggered, the final trained LSTM network was obtained, which is the BO-LSTM correction model, used for correcting precipitation forecast errors for different forecast periods. Finally, the BO-LSTM model was tested using the entire year of 2023 data as an independent test set. Figure 2 This is a schematic diagram showing the results before and after the correction of the BO-LSTM rainfall forecast in this invention.

[0044] Step 3: Construction of the Bayesian Optimization-Based BO-RF Correction Model Similarly, in step 2, the dataset is first preprocessed: missing values ​​and outlier noise points are removed, the time resolution is standardized to a daily scale, and the predicted sunshine hours are converted into quantitative values. Then, a random forest feature importance ranking method is used to select the top three most important core features from all forecast factors, which will serve as the input features for the subsequent model. Using these core features from each station on that day as input and the measured rainfall at each station on that day as the output target, a training sample set is constructed.

[0045] The RF hyperparameters were optimized using Bayesian optimization with the same parameter configuration as in step 2. The hyperparameter search range was as follows: number of decision trees 50 to 500 (integers), maximum depth 10 to 200 (integers), minimum number of leaf nodes 1 to 20 (integers), and the mtry policy was discretely selected from {'sqrt', 'log2', 'all'} (considered as categorical variables). The optimization objective was to minimize the root mean square error (RMSE) of 5-fold cross-validation. Iterative updates were performed using a Gaussian process surrogate model, and the algorithm automatically converged to the optimal hyperparameter combination after 50 evaluations.

[0046] Using optimal hyperparameters, the random forest model was retrained on the entire training set (2019-2022). A bagging ensemble strategy was employed, and out-of-bag error was used to monitor model performance. The resulting BO-RF correction model was used for error correction in medium- and long-term precipitation forecasts for different lead times. Figure 3 This is a schematic diagram showing the results of the BO-RF rainfall forecast correction before and after the present invention.

[0047] Step 4: Determine the critical forecast period for switching between the two models According to S300 of the present invention, the training set data are modeled and tested using the BO-LSTM model and BO-RF model trained above, and the precipitation forecast RMSE of the two models is calculated in batches for all forecast periods from 1 to 40 days.

[0048] The RMSE of the two models was calculated for each forecast period on the test set (2023 data). The results show that when the forecast period is 1 to 3 days, the RMSE of the BO-LSTM model is significantly lower than that of the BO-RF model; when the forecast period reaches 4 days or more, the RMSE of the BO-RF model begins to fall below that of the BO-LSTM model. Therefore, the forecast period corresponding to the first time the RMSE of the BO-LSTM model is greater than that of the BO-RF model is 3 days, i.e., the critical forecast period for model switching, Tc = 3 days. Thus, 1 to 3 days is considered a short-term forecast and should be corrected using the BO-LSTM model; 4 to 40 days is considered a medium- to long-term forecast and should be corrected using the BO-RF model.

[0049] Step 5: Time-Adaptive Correction According to S400 of the present invention, based on the constructed time-adaptive switching mechanism, for any numerical weather forecast data to be corrected, firstly, its forecast lead time t∈[1,3] days is obtained, then the BO-LSTM correction model is called for correction, and the correction value PLSTM(t) is output; if t∈[4,40] days, then the BO-RF correction model is called for correction, and the correction value PRF(t) is output.

[0050] In summary, the intelligent correction method for short-, medium-, and long-term precipitation forecast errors provided by this invention has shown excellent performance in practical applications in the Shandong Yellow River irrigation area and can be widely used in regional 1- to 40-day short-, medium-, and long-term precipitation forecast correction and drought and flood disaster early warning scenarios.

[0051] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0052] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for intelligent correction of short-, medium-, and long-term precipitation forecast errors, characterized in that, Includes the following steps: S100. Obtain historical precipitation forecast data and measured meteorological data for the forecast area, perform preprocessing and error feature analysis, and construct a training dataset; S200. Based on the training dataset, construct and optimize the time-adaptive precipitation forecast correction model, including the BO-LSTM correction model for short-term precipitation error correction and the BO-RF correction model for medium- to long-term precipitation error correction. S300, determine the critical forecast period for switching between the LSTM model and the RF correction model, and then construct a timely adaptive switching mechanism; S400: Obtain the original precipitation forecast data to be corrected for the forecast area, identify the forecast lead time, and automatically match the corresponding correction model through the time-adaptive switching mechanism to perform stratified correction on the short, medium and long-term precipitation forecast data to obtain the corrected precipitation forecast value.

2. The intelligent correction method for short-, medium-, and long-term precipitation forecast errors according to claim 1, characterized in that, S100 includes the following steps: S101. Obtain historical weather forecast products and corresponding measured meteorological data from multiple meteorological stations within the target forecast area; The historical weather forecast products include forecasted rainfall, forecasted maximum temperature, forecasted minimum temperature, and forecasted sunshine hours, while the measured meteorological data is the measured daily rainfall. S102. Clean the collected data, remove missing data, and perform unified spatiotemporal resolution matching to eliminate abnormal noise data. Convert the forecast sunshine hours into quantitative values ​​and construct a standardized historical rainfall forecast sample dataset. S103. Based on the standardized historical rainfall forecast sample dataset, perform error feature analysis and construct a training dataset; The error characteristic analysis includes exploring the spatiotemporal distribution patterns of precipitation forecast errors, statistically analyzing the underreporting and false alarm rates of forecast models under different precipitation intensity levels, and comparing the differences in precipitation forecast errors among different sub-regions within the forecast area.

3. The intelligent correction method for short-, medium-, and long-term precipitation forecast errors according to claim 1, characterized in that, S200 includes the following steps: S201. Perform Z-score standardization on the meteorological data in the training dataset; The meteorological data includes forecasted rainfall, forecasted maximum temperature, forecasted minimum temperature, forecasted sunshine hours, and measured rainfall. S202. Based on the standardized training dataset, construct the LSTM base model and the RF base model respectively; S203. Using a Bayesian optimization framework, with the minimization of the root mean square error of K-fold cross-validation as the objective function, a global adaptive search is performed on the hyperparameters of the LSTM base model and the RF base model, respectively. S204. During the iterative search process, the optimal hyperparameter combination is obtained based on the minimum cross-validation error of the Bayesian optimization output, thereby obtaining the BO-LSTM correction model for short-term precipitation error correction and the BO-RF correction model for medium- to long-term precipitation error correction.

4. The intelligent correction method for short-, medium-, and long-term precipitation forecast errors according to claim 3, characterized in that, In step S202, for the LSTM base model, the predicted rainfall, predicted minimum temperature and predicted sunshine hours in the standardized training dataset are used as input features, and the measured rainfall is used as the prediction target. For the RF basic model, the multi-dimensional meteorological data in the standardized training dataset is used as input features, and the measured rainfall is used as the prediction target; wherein, the multi-dimensional meteorological data includes the forecast rainfall, the forecast maximum temperature, the forecast maximum temperature and the forecast sunshine duration.

5. The intelligent correction method for short-, medium-, and long-term precipitation forecast errors according to claim 3, characterized in that, In step S202, during the construction of the LSTM base model, the Adam optimizer is used, initial hyperparameters are configured for iterative training, and an early stopping mechanism is introduced to prevent model overfitting. The hyperparameters in the LSTM base model include the learning rate, the number of hidden units, the Dropout rate, and the backtracking window. In the process of constructing the RF basic model, multiple decision trees are integrated by bootstrapping and initial hyperparameters are configured. The model structure is initially adjusted by combining out-of-bag error. The hyperparameters in the RF basic model include the number of decision trees, the maximum number of splits, the minimum number of leaf nodes, and the feature material strategy.

6. The intelligent correction method for short-, medium-, and long-term precipitation forecast errors according to claim 3, characterized in that, In step S203, for the LSTM base model and the RF base model, in the global adaptive search process using the Bayesian optimization framework, the search space is set as a continuous or discrete interval as a hyperparameter based on prior knowledge, and the posterior distribution is iteratively updated through a Gaussian process surrogate model.

7. The intelligent correction method for short-, medium-, and long-term precipitation forecast errors according to claim 3, characterized in that, In S204, the method for determining the BO-LSTM correction model is as follows: After obtaining the optimal hyperparameter combination, the data in the training dataset are divided into a training subset and a validation subset in chronological order. The LSTM base model is retrained on the training subset using the optimal hyperparameter combination, and overfitting is monitored using the validation subset. At the same time, L2 regularization and early stopping mechanism are applied to obtain the BO-LSTM correction model for short-term precipitation error correction. The method for determining the BO-RF correction model is as follows: The RF base model is retrained using the optimal hyperparameter combination and concentrated data from the training dataset, and the model stability is improved by combining the Bagging ensemble strategy, thus obtaining the BO-RF correction model for medium- to long-term precipitation error correction.

8. The intelligent correction method for short-, medium-, and long-term precipitation forecast errors according to claim 1, characterized in that, S300 includes the following steps: S301. Use the BO-LSTM correction model and the BO-RF correction model respectively to model and test the unified training dataset, and calculate the corresponding error indexes for all short, medium and long-term forecast periods in batches. S302. For the error indices of the BO-LSTM correction model and the BO-RF model under the same forecast period, the forecast duration corresponding to the first time when the error index of the BO-LSTM correction model is greater than the error index of the BO-RF correction model is taken as the critical forecast period for switching between the two models. S303. Based on the critical forecast period for dual-model switching, a time-adaptive switching mechanism is constructed, including: When the forecast period is less than the critical forecast period for switching between the two models, it is determined to be a short forecast period, and the BO-LSTM correction model is called to correct the precipitation forecast error. When the forecast period is longer than the critical forecast period for switching between the two models, it is determined to be a medium- to long-term forecast, and the BO-RF correction model is called to correct the precipitation forecast error.

9. The intelligent correction method for short-, medium-, and long-term precipitation forecast errors according to claim 8, characterized in that, The error index is the precipitation forecast RMSE.

Images