Search

CN-121682780-B - Climate prediction method, equipment and medium based on hot start and background constraint

CN121682780BCN 121682780 BCN121682780 BCN 121682780BCN-121682780-B

Abstract

The invention discloses a climate prediction method, equipment and medium based on hot start and background constraint. The method comprises the steps of training a self-encoder by using actual climate data of a first historical annual interval and then taking the self-encoder as an initial climate prediction model decoder, acquiring first climate data of the current annual and a second historical annual interval predicted by a specified climate prediction model, generating a sample pair of the month by taking the month and the first climate data of the number of months set before as input features and taking the actual climate data of the month as a label for each month in the second historical annual interval, training a climate prediction model trained by the last year by using a training sample library with minimum joint loss as a target, wherein the joint loss comprises precision constraint representing model prediction performance, climate background constraint and prediction scoring constraint, and the trained model is used for predicting the climate of the current annual. The invention realizes the continuous evolution and stable prediction of the prediction model and improves the convergence rate of the prediction model.

Inventors

  • ZHANG BOFEI
  • YU HAIPENG
  • LIU XIN

Assignees

  • 中国科学院西北生态环境资源研究院

Dates

Publication Date
20260512
Application Date
20260210

Claims (9)

  1. 1. A method of climate prediction based on hot start and context constraints, comprising: Training the self-encoder by using the actual climate data of the first historical annual interval to enable the self-encoder to learn the dependency relationship among stations, and taking the weight of the self-encoder after training as an initial weight or regularization term of an initial climate prediction model decoder; Acquiring first climate data of a current year and a second historical year interval predicted by a specified climate system model, wherein the first historical year interval is earlier than the second historical year interval; for each month in the second historical annual interval, taking the first climate data of the month and the preset number of months as input characteristics and the actual climate data of the month as a label to generate a sample pair of the month; Training a weather prediction model trained in the last year by using the training sample library with the minimum combined loss as a target, wherein the combined loss comprises precision constraint, weather background constraint and prediction score constraint for representing the prediction performance of the model; and inputting the month to be predicted and the first climate data of the previously set number of months into a trained climate prediction model for any month to be predicted in the current year, and outputting a climate prediction result of the month to be predicted.
  2. 2. The warm-start and context constraint based climate prediction method of claim 1, wherein the joint loss is: ; Wherein, the For the joint loss; Is a first super parameter; Is a second super parameter; And The order of magnitude used to control the accuracy constraint, the climate background constraint, and the predictive scoring constraint is the same; Characterizing an error between a predicted value and an actual value for the precision constraint; For the climate background constraint, whether the fluctuation range and the extreme value probability of the predicted value accord with the climate rules of the first historical annual interval and the second historical annual interval of the station or not is represented; and characterizing the proximity between the predicted value and the actual value for the prediction score constraint.
  3. 3. The method for climate prediction based on hot start and background constraints according to claim 1 or 2, the method is characterized in that the climate background constraint is as follows: ; or the climate context constraint is: ; Wherein, the Constraining the climate background; Is a KL divergence function; Is a Wasserstein distance function; The distribution of the current batch prediction results of the climate prediction model is obtained; a cumulative distribution or probability density of actual climate data for the first historical annual interval.
  4. 4. The method for climate prediction based on hot start and background constraints according to claim 1 or 2, the prediction score constraint is characterized in that: ; Wherein, the For the prediction scoring constraints described above, In order to observe the total number of sites, For the number of observation sites for which the predicted value range-to-average percentage is positive and negative as the actual value range-to-average percentage, The absolute value of the predicted value distance flat percentage and the actual value distance flat percentage are both The number of observation sites within an interval, The absolute value of the predicted value distance flat percentage and the actual value distance flat percentage are both The number of observation sites within an interval, Is that the absolute value of the predicted value is not less than 1 and the absolute value of the actual value is at The number of observation sites within an interval, For the first preset parameter(s), For the second preset parameter, the second preset parameter is set, A is a first coefficient, b is a second coefficient, and c is a third coefficient.
  5. 5. The method for predicting climate based on hot start and background constraint according to claim 1, wherein obtaining the first climate data of the current year and the second historical year interval predicted by the specified climate system model comprises: acquiring data of each set member in the current annual and second historical annual intervals predicted by the specified climate system model; for each year, carrying out aggregate average on the data of all the aggregation members in the year to obtain a climatic factor field of the year, carrying out interpolation resampling on the climatic factor field to a preset standard space grid, and generating first climatic data of the year.
  6. 6. The warm-start and context constraint based climate prediction method of claim 1, wherein the climate prediction model comprises: an input adaptation layer for receiving multi-channel sliding time window data input to the climate prediction model; The deep feature extraction module comprises a residual block and a maximum pooling layer which are alternately connected, and is used for extracting nonlinear space features of multi-channel sliding time window data to obtain a feature map; The spatial feature aggregation module comprises an adaptive average pooling layer and is used for compressing the feature map into a specified size; The decoder comprises a full-connection layer and a nonlinear regression layer which are alternately connected, the nonlinear regression layer adopts an activation function and a random deactivation function to carry out nonlinear regression processing, and the decoder is used for mapping the characteristic map with the specified size into the predicted value of each site.
  7. 7. The method for predicting climate based on hot start and background constraint of claim 1, further comprising outputting a climate prediction result of each month of the current year by the climate prediction model after training of the current year, and calculating a quarter climate prediction result according to the climate prediction result of the corresponding month.
  8. 8. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the hot start and context constraint based climate prediction method of any of claims 1-7 when the program is executed by the processor.
  9. 9. A non-transitory computer readable storage medium storing computer instructions for causing a computer to perform the hot start and context constraint based climate prediction method of any of claims 1-7.

Description

Climate prediction method, equipment and medium based on hot start and background constraint Technical Field The invention belongs to the field of artificial intelligence and climate prediction, and in particular relates to a climate prediction method, equipment and medium based on hot start and background constraint. Background Climate prediction is an important decision basis for disaster prevention and reduction, water resource management and agricultural production planning. The deep learning technology provides a new idea for climate prediction due to the strong nonlinear fitting capability. However, climate systems are non-stationary and for weather forecasting, data often has small sample characteristics, and conventional deep learning downscaling methods often have the following limitations. On the one hand, the model updating is lagged, namely, the model parameters of the traditional deep learning method are kept unchanged after the prediction annual training is finished, the model parameters cannot be self-evolved by using the newly added observation data each year, if the model is trained from the beginning each year (cold start), the calculation resource waste is caused, and the model is possibly trapped into an unstable local optimal solution due to the scarcity of the sample size. On the other hand, long-period climate rules are lacking, the existing model is often trained by only paired data in the last decades (for example, 2001-2020), and long-sequence observation data which are corresponding to the unpowered mode but contain precious climate state distribution information in the earlier period (for example, 1951-2000) are ignored, so that the model has insufficient prediction capability on extreme climate events (for example, hundred-year storm). On the other hand, the physical consistency is insufficient, namely, the simple pursuit of root mean square deviation (RMSE) minimization easily leads to the fact that the prediction result tends to be smooth, and the extreme distribution characteristics of elements such as precipitation and the like are lost. Therefore, there is a need for a method that can continuously integrate new yearly observations and ensure predictive model stability and timeliness through smooth parameter updates. Disclosure of Invention The invention mainly aims to provide a climate prediction method, equipment and medium based on hot start and background constraint, so as to overcome the defects of the prior art. The first aspect of the invention provides a hot start and background constraint-based climate prediction method, which comprises the steps of training a self-encoder by using actual climate data of a first historical annual interval, taking the weight of the self-encoder after training as an initial weight or regularization term of an initial climate prediction model decoder after learning a dependency relation among stations of the self-encoder, acquiring first climate data of a current annual interval predicted by a specified climate prediction model and a second historical annual interval, wherein the first historical annual interval is earlier than the second historical annual interval, taking the first climate data of the month and a preset number of months as input characteristics for each month in the second historical annual interval, taking the actual climate data of the month as a label, generating a sample pair of the month, combining the sample pair of each month to form a training sample library, training a climate prediction model of a previous year by using the training sample library, and the joint loss comprises precision constraint, climate background constraint and prediction score for representing model prediction performance, and outputting a climate result to be predicted in the current annual month and a climate result to be predicted in the first month to be predicted, and outputting the weather result to be predicted in the current month to be predicted. Preferably, the joint loss is: ; Wherein, the For the joint loss; Is a first super parameter; Is a second super parameter; And The order of magnitude used to control the accuracy constraint, the climate background constraint, and the predictive scoring constraint is the same; Characterizing an error between a predicted value and an actual value for the precision constraint; For the climate background constraint, whether the fluctuation range and the extreme value probability of the predicted value accord with the climate rules of the first historical annual interval and the second historical annual interval of the station or not is represented; and characterizing the proximity between the predicted value and the actual value for the prediction score constraint. Preferably, the climate context constraint is: ; or the climate context constraint is: ; Wherein, the Constraining the climate background; Is a KL divergence function; Is a Wasserstein distance function; The distribution