Flood forecasting precision improving method based on multi-model data fusion driving

Abstract

The invention discloses a flood forecasting precision improving method based on multi-model data fusion driving, relates to the technical field of meteorological numerical forecasting data fusion, and aims to solve the problems of uncertainty of a numerical mode of simulated atmosphere, defects of the mode and forecasting errors of the numerical mode. The method comprises the following steps: respectively collecting output data of different numerical weather forecast models, and normalizing the output data; selecting a weather forecast model group with representativeness and diversity to obtain a combined weather forecast model and an ensemble forecast sample; performance evaluation is carried out on the ensemble forecasting samples obtained through the evaluation indexes, and if the performance reaches a preset target, meteorological numerical forecasting data fusion of multi-model random combination is completed; and otherwise, adjusting the random combination strategy and the integration model, and carrying out iterative optimization until a predetermined target is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological numerical prediction data fusion, and specifically to a method for improving the accuracy of flood prediction driven by multi-model data fusion. Background Art

[0002] Numerical prediction, as the mainstream technology of weather forecasting, has made revolutionary progress in recent years. It has become an important part of meteorological forecasting operations and is constantly developing towards a more refined direction. Based on the actual situation of the atmosphere, it solves the equations of fluid mechanics and thermodynamics through a supercomputer to predict the future state of atmospheric motion. However, due to the complexity of meteorology, combined with the uncertainty of numerical models for simulating the atmosphere and the defects of the models themselves, the prediction errors of numerical models still exist to this day. Therefore, it is of great significance to correct numerical predictions to obtain more accurate results. Although the prediction accuracy of a single numerical model can be improved by adjusting parameters, selecting optimal prediction factors, etc. during the prediction process, there is uncertainty in a single model itself and it is difficult to adapt to all meteorological situations. Numerous studies have shown that combining multiple single prediction models to construct a fused prediction model can effectively utilize the advantages of different numerical models, thereby improving the reliability and accuracy of predictions. Summary of the Invention

[0003] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for improving the accuracy of flood prediction driven by multi-model data fusion, including the following steps: S1, respectively collect the output data of different numerical weather prediction models and normalize the output data; S2, select a representative and diverse group of weather prediction models, evaluate each weather prediction model in the group of weather prediction models according to performance indicators, determine the prediction ability of each weather prediction model, assign a weight to each weather prediction model, and linearly weight and combine each weather prediction model to obtain a combined weather prediction model; S3, according to the random combination strategy, use a multi-objective optimization algorithm to optimize the upper and lower limit intervals of the weights of the combined weather prediction model to obtain the optimized upper and lower limit intervals of the weights; S4, use the optimized upper and lower limits of the weights as decision variables, randomly generate multiple groups of weight values within the interval to obtain an ensemble prediction sample; S5, perform performance evaluation on the ensemble prediction sample obtained through evaluation indicators. If the performance reaches the predetermined target, the meteorological numerical prediction data fusion of multi-model random combination is completed; otherwise, adjust the random combination strategy and the integrated model, and perform iterative optimization until the predetermined target is reached.

[0004] Further, the steps of selecting a representative and diverse group of weather forecasting models, evaluating each weather forecasting model in the group of weather forecasting models according to performance indicators, determining the prediction ability of each weather forecasting model, assigning a weight to each weather forecasting model, and linearly weighting and combining each weather forecasting model to obtain a combined weather forecasting model include: Step S21: Select a representative and diverse group of weather forecasting models, evaluate each weather forecasting model in the group of weather forecasting models according to performance indicators, and determine the prediction ability of each weather forecasting model, where the performance evaluation indicator is one of the mean square error or the mean absolute error; Step S22: Design a weight assignment scheme according to the performance evaluation results of the models, use the inverse of the model performance indicator as the weight, and normalize the weight; Step S23: Linearly weight and combine the forecast results of each model using the normalized weight. If there are N models, and the weight of the th model is , and the linear combination result is F, then the final linear combination F is: .

[0005] Further, according to the random combination strategy, the upper and lower limit intervals of the weights of the combined weather forecasting model are optimized by using a multi-objective optimization algorithm, including: Taking the upper limit and of the random weight as decision variables, and taking the minimum average relative interval width, the maximum interval coverage rate, and the minimum root mean square error, which are common evaluation indicators for interval forecasting, as optimization objectives, a multi-objective optimization model for the combined weight is established. The objective function and optimization variables of the model are respectively: In the formula, is the total number of forecast samples in the verification period of the deterministic model; is the serial number of the forecast sample sequence in the verification period of the deterministic model; is the measured value; is the forecast value corresponding to the 95th percentile in the ensemble forecast sample; is the forecast value corresponding to the 5th percentile in the ensemble forecast sample; is a conditional function, which is valid only when the conditions in the parentheses are all satisfied , otherwise ; is the mean of the ensemble forecast sample; is the set of optimization variables, where ; is the objective function; is the average relative interval width; is the interval coverage rate; is the root mean square error.

[0006] Furthermore, taking the optimized upper and lower limits of the weights as decision variables, multiple groups of weights are randomly generated within the interval to obtain an ensemble prediction sample, including: Step S41: According to the determined upper and lower limit intervals of the combined weights, use the Monte Carlo method to generate random weights within the upper and lower limit intervals of the combined weights; if is the weight vector, where n is the total number of models, and each weight is generated as follows: 1. For each model , set the uniform distribution interval of the weight ; is the uniform distribution interval of the weight of model i; 2. Independently draw a random number from the uniform distribution as the initial weight; 3. Apply the normalization process so that the sum of all weights is equal to 1: ; where represents the weight of model i obeys the uniform distribution; represents the jth random number; Step S42: Re-weight and combine the prediction results of each model, apply the generated weight vector to the prediction results of each model. Let be the prediction result of the th model. The ensemble prediction F is calculated by the following weighted summation formula: Step S43: Repeat the generation process to generate ensemble members. Repeat Step S41, use the Monte Carlo method to generate new weight vectors multiple times. For each newly generated set of weights, calculate the weighted combined prediction result according to Step S42 to generate new ensemble members, and store all the generated ensemble members in a set to form a sample set of the ensemble prediction.

[0007] Furthermore, for the performance evaluation of the ensemble prediction sample obtained through the evaluation index, if the performance reaches the predetermined target, the data fusion of the multi-model random combination of meteorological numerical prediction is completed; otherwise, adjust the random combination strategy and the integrated model, and perform iterative optimization until the predetermined target is reached, including: Step S51: Determine the indicators for evaluating the performance of the ensemble prediction according to the mean square error, mean absolute error, and correlation coefficient, and obtain the performance evaluation result; Step S52: adjust the weight of each model according to the performance evaluation result, and re-evaluate and adjust the upper and lower limits of the weight through a multi-objective optimization algorithm; Step S53, regenerate the ensemble forecast and iterate, use the updated weights and model adjustments, re-execute step S4, generate a new ensemble forecast sample set, and perform performance evaluation on the new ensemble forecast sample set again. If the performance reaches the predetermined target, stop the iteration; otherwise, continue the optimization.

[0008] The beneficial effect of the present invention is that the meteorological numerical forecast data fusion method based on the random combination of multiple models integrates the prediction results of multiple models, and its core beneficial effect is that it significantly enhances the comprehensive performance of the forecast results. This method utilizes the different sensitivities and prediction advantages of different models to data, reduces the possible deviations of a single model, effectively balances the strengths and weaknesses of each model, thereby reducing the systematic deviations that may be introduced by a single model and improving the overall accuracy of the forecast. In addition, through weight distribution and random combination strategies, the estimation of forecast uncertainty is enhanced, and the complexity and variability of meteorological phenomena are better captured, making the forecast results not only more accurate but also more robust. The generation of ensemble forecasts not only provides an estimate of uncertainty, but also increases the diversity of forecasts, providing decision makers with a more comprehensive perspective and information, thereby playing a key role in disaster prevention and mitigation and climate adaptation. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a flowchart of a flood forecasting accuracy improvement method driven by multi-model data fusion; Figure 2 Schematic diagram of the implementation of a flood forecasting accuracy improvement method driven by multi-model data fusion. DETAILED DESCRIPTION

[0010] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0011] The features and performance of the present invention are further described in detail below in conjunction with the embodiments.

[0012] like Figure 1 As shown in FIG. 1 , a flood forecasting accuracy improvement method based on multi-model data fusion drive includes the following steps: S1, respectively collect the output data of different numerical weather forecast models and normalize the output data; S2. Select a representative and diverse group of weather forecasting models, evaluate each weather forecasting model in the group according to performance metrics, determine the prediction ability of each weather forecasting model, assign a weight to each weather forecasting model, and linearly weight and combine the weather forecasting models to obtain a combined weather forecasting model; S3. According to the random combination strategy, use a multi-objective optimization algorithm to optimize the upper and lower limit intervals of the weights of the combined weather forecasting model to obtain the optimized upper and lower limit intervals of the weights; S4. Use the optimized upper and lower limits of the weights as decision variables, randomly generate multiple groups of weight values within the interval to obtain an ensemble forecast sample; S5. Through the performance evaluation of the ensemble forecast sample obtained by the evaluation index, if the performance reaches the predetermined target, the meteorological numerical forecast data fusion of the multi-model random combination is completed; otherwise, adjust the random combination strategy and the integrated model, and perform iterative optimization until the predetermined target is reached.

[0013] Further, the selection of a representative and diverse group of weather forecasting models, evaluating each weather forecasting model in the group according to performance metrics, determining the prediction ability of each weather forecasting model, assigning a weight to each weather forecasting model, and linearly weight and combining the weather forecasting models to obtain a combined weather forecasting model includes: Step S21. Select a representative and diverse group of weather forecasting models, evaluate each weather forecasting model in the group according to performance metrics, and determine the prediction ability of each weather forecasting model, where the performance evaluation metric is one of the mean square error or the mean absolute error; Step S22. According to the performance evaluation results of the models, design a weight assignment scheme, use the inverse of the model performance metric as the weight, and normalize the weight; Step S23. Linearly weight and combine the forecast results of each model using the normalized weights. If there are N models, and the weight of the th model is , and the linear combination result is F, then the final linear combination F is: .

[0014] Further, the use of the random combination strategy to optimize the upper and lower limit intervals of the weights of the combined weather forecasting model using a multi-objective optimization algorithm to obtain the optimized upper and lower limit intervals of the weights includes: Taking the upper limit and of the random weights as decision variables, and taking the minimum average relative interval width, the maximum interval coverage rate, and the minimum root mean square error, which are common evaluation metrics for interval forecasting, as optimization objectives, establish a multi-objective optimization model for the combined weights. The objective function and optimization variables of the model are: In the formula, is the total number of forecast samples in the verification period of the deterministic model; is the sequence number of the forecast samples in the verification period of the deterministic model; is the measured value; is the forecast value corresponding to the 95% quantile in the ensemble forecast samples; is the forecast value corresponding to the 5% quantile in the ensemble forecast samples; is a conditional function, which is only valid when the conditions in the parentheses are all satisfied , otherwise ; is the mean of the ensemble forecast samples; is the set of optimization variables, where ; is the objective function; is the average relative interval width; is the interval coverage rate; is the root mean square error.

[0015] Furthermore, taking the optimized upper and lower limits of the weights as decision variables, multiple groups of weights are randomly generated within the interval to obtain ensemble forecast samples, including: Step S41: According to the determined upper and lower limit intervals of the combined weights, use the Monte Carlo method to generate random weights within the upper and lower limit intervals of the combined weights; if is the weight vector, where n is the total number of models, and each weight is generated as follows: 1. For each model , set the uniform distribution interval of the weight ; is the uniform distribution interval of the weight of model i; 2. Independently draw a random number from the uniform distribution as the initial weight; 3. Apply the normalization process, and the sum of all weights is equal to 1: ; where represents the weight of model i obeys the uniform distribution; represents the jth random number; Step S42: Re-weight and combine the forecast results of each model, apply the generated weight vector to the forecast results of each model. Let be the forecast result of the th model. The ensemble forecast F is calculated by the following weighted summation formula: Step S43: Repeat the generation process to generate set members. Repeat Step S41 and use the Monte Carlo method to generate new weight vectors multiple times. For each newly generated set of weights, calculate the forecast results of the weighted combination according to Step S42, generate new set members, and store all the generated set members in a set to form a sample set for ensemble forecasting.

[0016] Furthermore, for the performance evaluation of the ensemble forecasting sample set obtained through the evaluation index, if the performance reaches the predetermined target, the meteorological numerical forecast data fusion of multi-model random combination is completed; otherwise, adjust the random combination strategy and the ensemble model, and perform iterative optimization until the predetermined target is reached, including: Step S51: Determine the indicators for evaluating the performance of the ensemble forecast according to the mean square error, mean absolute error, and correlation coefficient, and obtain the performance evaluation result. Step S52: According to the performance evaluation result, adjust the weights of each model, and re-evaluate and adjust the upper and lower limits of the weights through a multi-objective optimization algorithm. Step S53: Regenerate the ensemble forecast and iterate. Use the updated weights and model adjustments, re-execute Step S4, generate a new sample set of ensemble forecasts, and perform performance evaluation on the new sample set of ensemble forecasts again. If the performance reaches the predetermined target, stop the iteration; otherwise, continue to optimize.

[0017] Next, the accompanying drawings in the embodiments of the present invention will be combined: Embodiment 1: Refer to Figure 2 , the present invention provides a method for improving the accuracy of flood forecasting driven by multi-model data fusion. The implementation process of the present invention includes the following steps: Step S101: Collect data and preprocess it. Collect the output data from different numerical weather prediction models and perform quality control on the data. The quality control includes performing data standardization or normalization to eliminate the dimensional and magnitude differences between different data sources. Step S102: Determine the weights and perform linear combination. According to the prediction ability results, assign weights to each model and perform linear weighted combination. The step of assigning weights to each model according to the prediction ability results and performing linear weighted combination includes: selecting a set of representative and diverse numerical models, evaluating the performance of each model on historical data, determining their prediction ability, assigning a weight to each model. For example, use the inverse of the mean square error (MSE) of historical forecasts as the weight, and perform linear weighted combination on multiple forecast models to obtain a combined forecast model. Step S103: Design a random combination strategy and optimize the upper and lower limits of the combination weights using a multi-objective optimization algorithm; Step S104: Generate an ensemble forecast. Using the upper and lower limits of the random weights as decision variables, randomly generate multiple sets of weights within the interval to obtain ensemble forecast samples; Step S105: Iterative optimization. According to the performance evaluation and real feedback, adjust the random combination strategy and the ensemble model for iterative optimization.

[0018] Example 2: Refer to Figure 2 , further, step S102 for determining weights and linear combination specifically includes the following steps: Step S1021: Model performance evaluation. Select a set of representative and diverse numerical models and evaluate the performance of each model on historical data to determine their prediction capabilities. The mean squared error (MSE) or mean absolute error (MAE) can be used as performance metrics to evaluate the performance of each model on historical data.

[0019] Step S1022: Determine the weight scheme and calculate the weights. According to the performance evaluation metrics of the models, design a weight allocation scheme, and then, based on the performance evaluation results of the models, design a weight allocation scheme. Use the inverse of the model performance metrics as weights and normalize the weights.

[0020] Step S1023: Linear combination. Use the normalized weights to perform a linear weighted combination of the forecast results of each model. If there are N models, and the weight of the -th model is , and the linear combination result is F, then the final linear combination F is: .

[0021] Example 3: Refer to Figure 2 , further, step S104 for generating an ensemble forecast specifically includes the following steps: Step S1041: Use the Monte Carlo method to determine the randomness of the weights. According to the upper and lower limits of the combination weights determined by the multi-objective optimization algorithm in step S103, use the Monte Carlo method to generate random weights within this interval. If is the weight vector, where n is the total number of models, and the generation process of each weight is as follows: For each model , set the uniform distribution interval of the weight .

[0022] Independently draw a random number from the uniform distribution as the initial weight.

[0023] Apply the normalization process to ensure that the sum of all weights equals 1: ; Step S1042: Re-weight and combine the prediction results of each model. For the generated weight vector , we apply it to the prediction results of each model. Let be the prediction result of the -th model. The ensemble prediction F can be calculated by the following weighted summation formula: .

[0024] Step S1043: Repeat the generation process to generate ensemble members. Repeat Step S1041 and use the Monte Carlo method to generate new weight vectors multiple times. For each newly generated set of weights, calculate the weighted combined prediction results according to Step S1042 to generate new ensemble members. Store all the generated ensemble members in a set to form a sample set of the ensemble prediction. By randomly selecting weights within the weight interval, the diversity of the ensemble members can be ensured, reflecting the uncertainty of the prediction results.

[0025] Example 4: Refer to Figure 2 , further, the iterative optimization of Step S105 specifically includes the following steps: Step S1051: Performance evaluation and feedback analysis. Determine the indicators for evaluating the performance of the ensemble prediction according to indicators such as the mean square error (MSE), mean absolute error (MAE), and correlation coefficient. And collect the feedback from users and domain experts to understand the performance and requirements of the ensemble prediction in practical applications.

[0026] Step S1052: Strategy adjustment and model update. According to the performance evaluation results, the weights of each model can also be adjusted. Use the multi-objective optimization algorithm to re-evaluate and adjust the upper and lower limits of the weights to improve the performance of the ensemble prediction.

[0027] Step S1053: Regenerate the ensemble prediction and iterate. Use the updated weights and possible model adjustments to re-execute Step S104 to generate a new sample set of the ensemble prediction. Perform performance evaluation on the new sample set of the ensemble prediction again. If the performance reaches the predetermined goal, stop the iteration; otherwise, continue to optimize. Self-correction based on the actual observed data and user feedback can gradually improve the accuracy and reliability of the ensemble prediction.

Claims

1. A flood forecasting accuracy improvement method based on multi-model data fusion drive, characterized in that: The steps include: S1, respectively collect the output data of different numerical weather forecast models and normalize the output data; S2, selecting a representative and diverse weather forecast model group, and evaluating each weather forecast model in the weather forecast model group according to the performance index, determining the prediction ability of each weather forecast model, assigning a weight to each weather forecast model, and performing a linear weighted combination of each weather forecast model to obtain a combined weather forecast model; S3, according to the random combination strategy, a multi-objective optimization algorithm is used to optimize the upper and lower limits of the weight of the combined weather forecast model to obtain the optimized upper and lower limits of the weight; S4, using the optimized upper and lower limits of the weight as decision variables, randomly generating multiple sets of weights within the interval to obtain ensemble forecast samples; S5, the performance of the ensemble forecast samples is evaluated by the evaluation index. If the performance reaches the predetermined target, the fusion of meteorological numerical forecast data of the random combination of multiple models is completed; otherwise, the random combination strategy and the integrated model are adjusted and iterative optimization is performed until the predetermined target is achieved.

2. The method for improving flood forecasting accuracy based on multi-model data fusion drive according to claim 1 is characterized in that: The method of selecting a representative and diverse weather forecast model group, evaluating each weather forecast model in the weather forecast model group according to performance indicators, determining the prediction ability of each weather forecast model, assigning a weight to each weather forecast model, and performing linear weighted combination of each weather forecast model to obtain a combined weather forecast model includes: Step S21, selecting a weather forecast model group that is representative and diverse, and evaluating each weather forecast model in the weather forecast model group according to a performance index to determine the prediction capability of each weather forecast model, wherein the performance evaluation index is one of a mean square error or a mean absolute error; Step S22: design a weight distribution scheme based on the performance evaluation results of the model, use the inverse of the model performance index as the weight, and normalize the weight; Step S23: Use the normalized weights to perform a linear weighted combination on the forecast results of each model. If there are N models, The weight of the model is , the linear combination result is F, then the final linear combination F is: 。 3. The method for improving flood forecasting accuracy based on multi-model data fusion drive according to claim 2 is characterized in that: The above-mentioned random combination strategy uses a multi-objective optimization algorithm to optimize the upper and lower limits of the weight of the combined weather forecast model to obtain the optimized upper and lower limits of the weight, including: With an upper limit of random weights and As the decision variable, the optimization objectives are the minimum average relative interval width, the maximum interval coverage rate and the minimum root mean square error of the commonly used evaluation indicators of interval forecasting. A multi-objective optimization model with combined weights is established. The objective function and optimization variables of the model are: In the formula, is the total number of forecast samples in the validation period of the deterministic model; Number the forecast sample sequence for the deterministic model validation period; is the measured value; is the forecast value corresponding to the 95% quantile in the ensemble forecast sample; is the forecast value corresponding to the 5% quantile in the ensemble forecast sample; It is a conditional function. It is valid only when the conditions in the brackets are met at the same time. ,otherwise ; is the mean of the ensemble forecast samples; is the set of optimization variables, where ; is the objective function; is the average relative interval width; is the interval coverage rate; is the root mean square error.

4. The method for improving flood forecasting accuracy based on multi-model data fusion drive according to claim 3 is characterized in that: The above method uses the optimized upper and lower limits of the weight as decision variables, randomly generates multiple sets of weights within the interval, and obtains ensemble forecast samples, including: Step S41: Generate random weights in the upper and lower limits of the combined weights using the Monte Carlo method according to the determined upper and lower limits of the combined weights; if is a weight vector, where n is the total number of models and each weight The generation process includes: 1) For each model , set the uniform distribution interval of weights ; is the uniform distribution interval of the weight of model i; 2) Independently from a uniform distribution Extract random numbers from As initial weight; 3) Apply the normalization process so that the sum of all weights equals 1: ;in Represents the weight of model i Follow uniform distribution; represents the jth random number; Step S42: weight the forecast results of each model again and generate a weight vector , applied to the forecast results of each model, assuming For the The forecast results of each model, the ensemble forecast F is calculated by the following weighted summation formula: ； Step S43, repeat the generation process to generate set members, repeat step S41, and use the Monte Carlo method to generate new weight vectors multiple times For each group of newly generated weights, the forecast result of the weighted combination is calculated according to step S42 to generate new set members, and all the generated set members are stored in a set to form a sample set of the set forecast.

5. The method for improving flood forecasting accuracy based on multi-model data fusion drive according to claim 4 is characterized in that: The performance evaluation of the ensemble forecast samples obtained by the evaluation index is described above. If the performance reaches the predetermined target, the fusion of meteorological numerical forecast data of the random combination of multiple models is completed; Otherwise, adjust the random combination strategy and integrated model and perform iterative optimization until the predetermined goal is achieved, including: Step S51, determining an indicator for evaluating the ensemble forecast performance according to the mean square error, the mean absolute error, and the correlation coefficient, and obtaining a performance evaluation result; Step S52: adjust the weight of each model according to the performance evaluation result, and re-evaluate and adjust the upper and lower limits of the weight through a multi-objective optimization algorithm; Step S53, regenerate the ensemble forecast and iterate, use the updated weights and model adjustments, re-execute step S4, generate a new ensemble forecast sample set, and perform performance evaluation on the new ensemble forecast sample set again. If the performance reaches the predetermined target, stop the iteration; otherwise, continue the optimization.

Images