A regional short-term temperature intelligent forecasting method based on process interpretable analysis

Abstract

This paper discloses a method for intelligently forecasting regional short-term temperature based on process-interpretable analysis. This method proposes process-interpretable analysis of temperature forecasts to reveal the key factors influencing these forecasts. By retraining the regional temperature forecast model and using contribution rates as feedback to the model input, this method enhances understanding of the temperature forecast process and provides effective feedback to improve the model's performance, ultimately leading to more accurate regional temperature forecasts.

Description

Technical Field

[0001] The present invention belongs to the technical field of temperature prediction, and in particular relates to a method for intelligently predicting regional short-term temperature based on process-interpretable analysis. Background Art

[0002] In the field of meteorology, current mainstream temperature forecasts can be divided into two categories based on the time span: long-term and short-term forecasts. Each category can be further divided into single-variable and multivariable forecasts. Current temperature forecasting methods can also be divided into numerical and statistical methods. Statistical methods can be further divided into traditional and modern statistical methods. Although current research has achieved some success in the field of temperature forecasting, several challenges remain.

[0003] Most current temperature prediction methods use numerical methods, which combine mathematics and physics to establish equations to predict or estimate / invert the temperature. The main drawback is that numerical methods cannot be based on historical data, making it difficult to determine parameters and solve equations. Statistical methods build models based on observed data, but when the amount of data is small, traditional statistical methods cannot achieve ideal prediction accuracy. Most literature or research using modern statistical methods almost all use a single machine learning model, so the accuracy is not high;

[0004] In the research based on temperature prediction, most of the above existing methods lack the research on the interpretability of the model; although some documents conduct interpretability analysis, they mainly use tree-type shallow interpreters (TreeExplainer), and still lack the "process" interpretable analysis of time series data. At the same time, the characteristic variables used for temperature prediction in most documents are single variables. For example, only using the temperature single variable for prediction often does not result in high prediction accuracy; even if some documents or studies use multiple characteristic variables, they only consider other meteorological variables (such as air pressure, relative humidity, etc.) other than the temperature of a single station (single meteorological detection station), and do not consider the data of other meteorological detection stations, which will also make the prediction accuracy low. The characteristic variables of some references use single-step data with a lag time of 1 hour, while the patent of this invention uses "multi-step" data with a lag time of 8 hours that comprehensively considers the prediction accuracy, computing resources and background knowledge of weather processes;

[0005] In addition, most literature or research uses the loss function that comes with artificial intelligence, and therefore does not consider the impact of some "extreme values" in the temperature time series on the prediction results. As a result, the prediction results will be affected by the extreme values ​​of "small samples" (that is, the sample values ​​are relatively small in number and the sample size is relatively small compared to the normal values), so the prediction accuracy is not high. Summary of the Invention

[0006] In order to solve the problems existing in the prior art, the present invention proposes a regional short-term temperature intelligent prediction method based on process interpretable analysis.

[0007] The technical solutions of the present invention are as follows:

[0008] A method for intelligent prediction of regional short-term temperature based on process-interpretable analysis, wherein the region is provided with a target station and a reference station, including:

[0009] Obtaining the temperature time series and auxiliary variable time series of the target station and the temperature time series of the reference station;

[0010] Based on the similarity between the temperature time series of the target station and the reference station, similar stations are screened among the reference stations while considering the nonlinear relationship between the temperature time series and the temperature time series;

[0011] Based on the temperature time series of the selected similar stations and the temperature time series and auxiliary variable time series of the target station, the characteristic value samples are constructed with the set lag time;

[0012] Based on the characteristic value samples, a pre-built regional temperature prediction model is initially trained, and an initial regional temperature prediction result is obtained based on the regional temperature prediction model that has been initially trained;

[0013] Performing a process interpretable analysis on the initial regional temperature prediction results to obtain process interpretable contribution rates of different eigenvalue samples for measuring the importance of the eigenvalue samples to the temperature prediction; based on the process interpretable contribution rates, updating the regional temperature prediction model that has completed the initial training to highlight the importance of the salient feature samples to the temperature prediction;

[0014] Based on the characteristic value samples, the updated regional temperature prediction model is trained a second time, and based on the regional temperature prediction model after the second training, a secondary regional short-term temperature prediction result is obtained and used as the final regional short-term temperature intelligent prediction result.

[0015] Furthermore, the auxiliary variables include air pressure, relative humidity, water vapor pressure, wet bulb temperature, wind speed and daily temperature variation time.

[0016] Furthermore, the specific steps of obtaining the temperature time series and auxiliary variable time series of the target station and the temperature time series of the reference station include:

[0017] Obtaining historical meteorological characteristic values ​​of the target station and historical temperature values ​​of the reference station, wherein the meteorological characteristics include temperature and auxiliary variables;

[0018] The historical meteorological characteristic values ​​of the target station and the historical temperature values ​​of the reference station are respectively subjected to preprocessing including missing value interpolation processing and feature normalization processing, and sorting processing by timestamp to form the temperature time series and auxiliary variable time series of the target station and the temperature time series of the reference station; wherein,

[0019] The specific method of the missing value interpolation process is: using the K-nearest neighbor algorithm for interpolation;

[0020] The specific method of the feature normalization processing is: using the Max-Min normalization method to perform feature normalization processing on historical meteorological feature values ​​and historical temperature values. The specific processing formula is as follows:

[0021]

[0022] Where y is the historical meteorological characteristic value or historical temperature value after feature normalization; x is the original value of the historical meteorological characteristic value or historical temperature value; x min is the minimum value among the original values ​​of historical meteorological characteristic values ​​or historical temperature values; x max It is the maximum value among the original values ​​of historical meteorological characteristic values ​​or historical temperature values.

[0023] Furthermore, the specific method for screening similar stations in the reference station based on the similarity between the temperature time series of the target station and the reference station while considering the nonlinear relationship between the temperature time series and the temperature time series includes:

[0024] Calculate the similarity between the temperature time series of the target station and the reference station, and select several reference stations with the highest similarity to the target station as similar stations; the similarity calculation formula is as follows:

[0025]

[0026] where p(x1) is the marginal distribution of the temperature time series x1 of the target station; p(y1) is the marginal distribution of the temperature time series y1 of the reference station; p(x1,y1) is the joint distribution of the temperature time series x1 of the target station and the temperature time series y1 of the reference station; I(x1;y1) is the similarity between the temperature time series x1 of the target station and the temperature time series y1 of the reference station, that is, the relative entropy of the joint distribution and marginal distribution between the temperature time series x1 of the target station and the temperature time series y1 of the reference station.

[0027] Furthermore, the specific method of constructing the characteristic value sample with the set lag time based on the temperature time series of the screened similar stations and the temperature time series and auxiliary variable time series of the target station includes:

[0028] Set the lag time n that takes into account both prediction accuracy and task requirements;

[0029] Based on the temperature time series of the selected similar stations, the temperature time series of the target station, and the auxiliary variable time series, a sample with characteristic values ​​is constructed:

[0030] Θ={X1,X2,...,X m-1 ,X m}

[0031] X1={x1(hn),x1(h-n+1),...,x1(h)}

[0032] X2={x2(hn),x2(h-n+1),...,x2(h)} ...

[0034] X m-k ={x m-k (hn),x m-k (h-n+1),...,x m-k (h)}

[0035] X m-k+1 ={x m-k+1 (hn),x m-k+1 (h-n+1),...,x m-k+1 (h)} ...

[0037] X m ={x m (hn),x m (h-n+1),...,x m (h)}

[0038] Where Θ is the eigenvalue sample set; X1 is the first subset formed based on the temperature time series of the target station; X2~X m-k X is the 2nd to mkth subsets formed based on the auxiliary variable time series of the target station; m-k+1 ~X m The m-k+1 to m-th subsets are formed based on the temperature time series of similar stations, and m is the number of variable dimensions of the eigenvalue sample;

[0039] x1(hn), x1(h-n+1), ..., x1(h) are the eigenvalue samples of the temperature of the target station corresponding to time hn, h-n+1, ​​..., h, respectively, and n is the number of lag time dimensions of the eigenvalue samples; x2(hn), x2(h-n+1), ..., x2(h) are the eigenvalue samples of the first auxiliary variable of the target station corresponding to time hn, h-n+1, ​​..., h, respectively;

[0040] x m-k (hn),x m-1 (h-n+1),...,x m-1 (h) are the characteristic value samples of the m-2th auxiliary variable of the target station corresponding to time hn, h-n+1, ​​..., h, and the total number of auxiliary variables is m-2;

[0041] x m-k+1 (hn),x m-k+1 (h-n+1),...,x m-k+1 (h) are the characteristic value samples of the temperature of the first similar station corresponding to time hn, h-n+1,…,h respectively;

[0042] x m (hn),x m (h-n+1),...,x m (h) are the characteristic value samples of the temperature at the k-1th similar station corresponding to time hn, h-n+1, ​​..., h, and the total number of similar stations is k-1;

[0043] Then the relationship between the eigenvalue sample and the actual temperature is:

[0044]

[0045] Where y(h+k) is the actual temperature value at the prediction time h+k; f(*) is the relationship model between the eigenvalue sample and the actual temperature; ε is the error.

[0046] Furthermore, the temperature prediction model adopts a nonlinear convex programming with constraints to integrate a deep foundation model;

[0047] The objective function of the temperature prediction model is:

[0048]

[0049] Where Z is the difference between the predicted temperature and the actual temperature; α is the α-norm; w j is the integration weight of the deep base model j to the minimization function, j∈[1,k], k is the total number of deep base models; y i is the actual temperature value of the predicted value i; is the temperature prediction value of deep base model j for the predicted value i, n is the total number of values ​​to be predicted; λ is the regularization parameter used to balance the two objective terms of the minimization function; w j The norm of

[0050] The constraints of the temperature prediction model are:

[0051] The loss function of the basic model in the temperature prediction model is:

[0052]

[0053]

[0054] Where ρ(r) is the loss function value corresponding to the basic model error r; r is the difference between the temperature prediction value of the basic model and the actual temperature value, that is, the basic model error; c is the robust scaling constant.

[0055] Furthermore, the deep basic model integrated in the temperature prediction model includes a gated recurrent unit, a bidirectional gated recurrent unit convolutional neural network, a long short-term memory network convolutional neural network coupled with a long short-term memory network, a convolutional neural network coupled with a bidirectional long short-term memory network, a convolutional neural network coupled with a gated recurrent unit, and a convolutional neural network coupled with a bidirectional gated recurrent unit.

[0056] Furthermore, the temperature prediction model uses the standard deviation index SD, the root mean square error index RMSE and the correlation coefficient index CC to evaluate the model prediction accuracy. The expressions of the standard deviation index SD, the root mean square error index RMSE and the correlation coefficient index CC are as follows:

[0057]

[0058]

[0059]

[0060] Where x i is the difference between the predicted temperature value and the actual temperature value of the predicted value i; i is the actual temperature value of the predicted value i; is the mean of the actual temperature value; P i is the predicted temperature value of the predicted value i; μ is the mean of the difference between the predicted temperature value of the predicted value i and the actual temperature value, n is the number of values ​​to be predicted.

[0061] Furthermore, the process interpretable analysis of the initial regional temperature forecast result is performed to obtain a specific method for the process interpretable contribution rate of different eigenvalue samples for measuring the importance of the eigenvalue samples to the temperature forecast:

[0062] Based on the relationship between the meteorological characteristics and the lag time in the characteristic value samples, a set Δ of variable and time step combinations is formed,

[0063] Δ={(i,t):1≤i≤M,1≤t≤N}

[0064] Where (i, t) is the combination of the i-th variable and the t-th time step; M is the total number of variables, corresponding to the number of variable dimensions in the eigenvalue sample; N is the total number of time steps, corresponding to the number of lag time dimensions in the eigenvalue sample;

[0065] Calculate the ProcessSHAP value for different combinations of variables and time steps. The ProcessSHAP value P for the combination of the i-th variable and the t-th time step is (i,t) ,

[0066]

[0067]

[0068] Where, is the characteristic function, which means that when the input X * The predicted output when there are only variable time pairs in the set S; Δ i is the set of all time steps of variable i; set S is a subset of all features; x S is a subvector of x, representing the features in the set S; f(·) represents the prediction model; x represents the input of the prediction model;

[0069] The sum of ProcessSHAP values ​​f(x * ),

[0070]

[0071] Where, P 0,0 This is the output of the feature function when all features do not exist. The sum of the feature function outputs for all combined feature sets;

[0072] Normalize the ProcessSHAP values ​​under different combinations of variables and time steps to obtain the process explanation contribution rate of different combinations of variables and time steps. For the combination of the i-th variable and the t-th time step, the process explanation contribution rate C (i,t) The calculation formula is as follows:

[0073]

[0074] The process based on the combination of variables and time steps can explain the contribution rate, and the process of obtaining different eigenvalue samples to measure the importance of eigenvalue samples to temperature prediction can explain the contribution rate.

[0075] Furthermore, the specific steps of updating the importance of salient feature samples to temperature prediction in the regional temperature prediction model that has been initially trained based on the explainable contribution rate of the process include:

[0076] The process explainable contribution rate is coupled as a weight to the convolution kernel of the first convolution layer in the regional temperature prediction model that has been initially trained, thereby updating the importance of the salient feature samples to the temperature prediction. The coupled convolution layer is calculated as follows:

[0077]

[0078] Where x is the xth row in the data matrix, y is the yth column in the data matrix, the operator * represents the convolution operation; l is the number of convolution layers; θ (l) is the activation function of the lth convolutional layer; w (l) is the convolution kernel of the lth convolution layer; A×B is the receptive field size of the convolution kernel, and z is the sliding step size; is the bias term of the lth convolutional layer; C (i,t) is the process-interpretable contribution of the eigenvalue sample corresponding to the combination of the i-th variable and the t-th time step.

[0079] Furthermore, in the secondary training, the model parameters of all convolutional layers are fixed, and only the model parameters of one fully connected layer are adjusted.

[0080] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor is configured to call and run the computer program stored in the memory to execute any of the above methods.

[0081] A computer-readable storage medium stores a computer program, wherein the computer program implements the steps of any of the above methods when executed by a processor.

[0082] Compared with the prior art, the present invention has the following beneficial effects:

[0083] The present invention proposes a regional short-term temperature intelligent prediction method based on process interpretable analysis. The method is based on the similarity between the temperature time series of the target station and the reference station, screens similar stations among the reference stations, and constructs characteristic value samples and trains the temperature prediction model based on the temperature time series of similar stations and the temperature time series and auxiliary variable time series of the target station. The design adopts "multivariables" with spatial information to explore the relationship between different variables, and establishes a temperature prediction model with a "spatiotemporal multivariable multi-step (lag time) integrated deep basic model", which improves the prediction accuracy.

[0084] Compared with similar station screening using Euclidean distance or Pearson correlation coefficient, the method of the present invention considers from the mutual information dimension, takes into account the nonlinear relationship of data between temperature time series and within temperature time series, and screens similar stations. In this way, similar stations with higher similarity to the target station can be screened out, and the changing trend of temperature can be explored, thereby improving the accuracy of model prediction to a certain extent.

[0085] Different from interpretability analysis using tree-based shallow interpreters, this method, based on Deep SHAP analysis of temperature forecast results, proposes a "process-based" interpretability analysis of temperature forecasts to reveal the key factors influencing process-based temperature predictions. This paves the way for the present invention's secondary training regional temperature prediction model.

[0086] The present invention conducts secondary training on the regional temperature prediction model, performs initial training based on the constructed feature samples, and performs process interpretable analysis on the initial regional temperature prediction results to obtain a process interpretable contribution rate for measuring the importance of eigenvalue samples to temperature prediction. Based on the process interpretable contribution rate, the regional temperature prediction model that has completed the initial training is updated to highlight the importance of eigenvalue samples to temperature prediction, guiding the prediction model to focus on learning this section (or a certain process) of data, thereby constructing a regional temperature prediction model that is more "focused" on these eigenvalue samples. This update effectively shifts the focus of the prediction model in the convolution process. The present invention uses the contribution rate as feedback to the input part of the model, enhances the understanding of the temperature prediction process, and provides effective feedback for the model, in order to improve its performance, thereby obtaining more accurate regional temperature prediction results.

[0087] During the secondary training process, the present invention fixes the model parameters of all convolutional layers and adjusts only the model parameters of the last fully connected layer. This design ensures that only a small number of parameters are optimized during the secondary training process, thereby improving the timeliness of temperature prediction and better facilitating the practical application of the results. The regression model obtained through this training process effectively improves the performance of temperature prediction.

[0088] The present invention adopts a set lag time when constructing the characteristic value sample. The lag time can be set according to the prediction accuracy, task requirements and meteorological background of the weather process, which is convenient for technicians to choose according to their own situation (professional field or research field).

[0089] The present invention constructs a temperature prediction model based on a nonlinear convex programming with constraints and an integrated deep basic model framework. The design of integrated learning can obtain more information at the bottom of the data and obtain better prediction results.

[0090] This method differs from loss functions commonly used in artificial intelligence models, such as mean absolute error (MAE) and mean square error (MSE). This method addresses the extreme values ​​inherent in meteorological data by using robust statistical functions such as the M-estimator (Cauchy and modified Huber functions) as a loss function. This reduces the impact of extreme values ​​in long temperature series on forecast results and improves temperature forecast accuracy. The robust scaling constant in the M-estimator can be customized for different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 Schematic diagram of the process of the regional short-term temperature intelligent prediction method based on process interpretable analysis in the embodiment;

[0092] Figure 2 2 is a principle block diagram of a method for intelligent prediction of regional short-term temperature based on process-interpretable analysis in an embodiment;

[0093] FIG3( a ) is a distribution map of national-level sites in Anhui Province in an application example;

[0094] FIG3( b ) is a distribution diagram of representative stations in an application embodiment;

[0095] Figure 4 This is a root mean square error analysis diagram of temperature prediction for different time periods using different loss functions in the application embodiment;

[0096] Figure 5(a) is a schematic diagram of the use of standard deviation values ​​to measure the accuracy of temperature prediction for different durations by different models at station number 58112 in an application embodiment; Figure 5(b) is a schematic diagram of the use of standard deviation values ​​to measure the accuracy of temperature prediction for different durations by different models at station number 58118 in an application embodiment; Figure 5(c) is a schematic diagram of the use of standard deviation values ​​to measure the accuracy of temperature prediction for different durations by different models at station number 58321 in an application embodiment; Figure 5(d) is a schematic diagram of the use of standard deviation values ​​to measure the accuracy of temperature prediction for different durations by different models at station number 58423 in an application embodiment;

[0097] Figure 6(a) is a scatter plot of the predicted and observed values ​​of the temperature for one hour corresponding to the test set in the application embodiment; Figure 6(b) is a scatter plot of the predicted and observed values ​​of the temperature for three hours corresponding to the test set in the application embodiment; Figure 6(c) is a scatter plot of the predicted and observed values ​​of the temperature for six hours corresponding to the test set in the application embodiment; Figure 6(d) is a scatter plot of the predicted and observed values ​​of the temperature for twelve hours corresponding to the test set in the application embodiment;

[0098] Figure 7 This is a scatter plot of the process characteristic density of temperature predictions at different stations based on the SHAP deep interpreter in the application example.

[0099] FIG8(a) is a time series diagram of the predicted and observed values ​​of the temperature corresponding to station number 58112 1 hour, 2 hours, and 3 hours in advance in an application embodiment; FIG8(b) is a time series diagram of the predicted and observed values ​​of the temperature corresponding to station number 58118 1 hour, 2 hours, and 3 hours in advance in an application embodiment; FIG8(c) is a time series diagram of the predicted and observed values ​​of the temperature corresponding to station number 58321 1 hour, 2 hours, and 3 hours in advance in an application embodiment; FIG8(d) is a time series diagram of the predicted and observed values ​​of the temperature corresponding to station number 58423 1 hour, 2 hours, and 3 hours in advance in an application embodiment;

[0100] Figure 9 This is a heatmap analysis diagram of the process of temperature prediction at different stations in the application example. DETAILED DESCRIPTION

[0101] The present invention is further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art fall within the scope defined by the claims attached to this application.

[0102] Example 1:

[0103] The present invention provides a regional short-term temperature intelligent prediction method based on process interpretable analysis. The region is equipped with several meteorological detection stations including target stations and reference stations, such as Figure 1 and Figure 2 Shown, including:

[0104] Obtain the temperature time series and auxiliary variable time series of the target station and the temperature time series of the reference station;

[0105] Based on the similarity between the temperature time series of the target station and the reference station, similar stations are screened among the reference stations, taking into account the nonlinear relationship between the temperature time series and the data within the temperature time series.

[0106] Based on the temperature time series of the selected similar stations and the temperature time series and auxiliary variable time series of the target station, the characteristic value samples are constructed with the set lag time;

[0107] Based on the characteristic value samples, the pre-built regional temperature prediction model is initially trained, and the initial regional temperature prediction result is obtained based on the regional temperature prediction model completed by the initial training;

[0108] A process-interpretable analysis was conducted on the initial regional temperature forecast results to obtain the process-interpretable contribution rate of different eigenvalue samples, which is used to measure the importance of eigenvalue samples to temperature prediction. Based on the process-interpretable contribution rate, the regional temperature prediction model that completed the initial training was updated to determine the importance of salient feature samples to temperature prediction.

[0109] Based on the eigenvalue samples, the updated regional temperature prediction model is trained a second time. Based on the regional temperature prediction model completed by the secondary training, the secondary regional short-term temperature prediction results are obtained and used as the final regional short-term temperature intelligent prediction results.

[0110] Example 2:

[0111] This embodiment is further designed based on the first embodiment in that the auxiliary variables in this embodiment include air pressure, relative humidity, water vapor pressure, wet-bulb temperature, wind speed and daily temperature variation time.

[0112] Example 3:

[0113] This embodiment is further designed based on the first embodiment in that the specific steps of obtaining the temperature time series and auxiliary variable time series of the target station and the temperature time series of the reference station in this embodiment include:

[0114] Obtain historical meteorological characteristic values ​​of the target station and historical temperature values ​​of the reference station. Meteorological characteristics include temperature and auxiliary variables.

[0115] The historical meteorological characteristic values ​​of the target station and the historical temperature values ​​of the reference station are preprocessed, including missing value interpolation and feature normalization, and sorted by timestamp, to form the temperature time series and auxiliary variable time series of the target station and the temperature time series of the reference station;

[0116] The specific method for interpolation of missing values ​​is as follows:

[0117] The K-Nearest Neighbors (KNN) algorithm is used for interpolation. The principle of KNN is to find the "closest" K samples in the data set and use the average of the K samples for interpolation;

[0118] The specific method of feature normalization is:

[0119] The Max-Min normalization method is used to perform feature normalization on historical meteorological characteristic values ​​and historical temperature values. The specific processing formula is as follows:

[0120]

[0121] Where y is the historical meteorological characteristic value or historical temperature value after feature normalization; x is the original value of the historical meteorological characteristic value or historical temperature value; x min is the minimum value among the original values ​​of historical meteorological characteristic values ​​or historical temperature values; x max It is the maximum value among the original values ​​of historical meteorological characteristic values ​​or historical temperature values.

[0122] Example 4:

[0123] This embodiment is further designed based on the first embodiment. In this embodiment, based on the similarity between the temperature time series of the target station and the reference station, a specific method for screening similar stations in the reference station by considering the nonlinear relationship between the temperature time series and the data within the temperature time series includes:

[0124] Calculate the similarity between the temperature time series of the target station and the reference station, and select the reference stations with the highest similarity to the target station as similar stations. The similarity calculation formula is as follows:

[0125]

[0126] where p(x1) is the marginal distribution of the temperature time series x1 of the target station; p(y1) is the marginal distribution of the temperature time series y1 of the reference station; p(x1,y1) is the joint distribution of the temperature time series x1 of the target station and the temperature time series y1 of the reference station; I(x1;y1) is the similarity between the temperature time series x1 of the target station and the temperature time series y1 of the reference station, that is, the relative entropy of the joint distribution and marginal distribution between the temperature time series x1 of the target station and the temperature time series y1 of the reference station.

[0127] Embodiment 5:

[0128] This embodiment is further designed based on the first embodiment. In this embodiment, based on the temperature time series of the screened similar stations and the temperature time series and auxiliary variable time series of the target station, a specific method for constructing a characteristic value sample with a set lag time includes:

[0129] Set the lag time n that takes into account both prediction accuracy and task requirements;

[0130] Construct eigenvalue samples based on the temperature time series of the selected similar stations, the temperature time series of the target station, and the auxiliary variable time series:

[0131] Θ={X1,X2,...,X m-1 ,X m}

[0132] X1={x1(hn),x1(h-n+1),...,x1(h)}

[0133] X2={x2(hn),x2(h-n+1),...,x2(h)} ...

[0135] X m-k ={x m-k (hn),x m-k (h-n+1),...,x m-k (h)}

[0136] X m-k+1 ={x m-k+1 (hn),x m-k+1 (h-n+1),...,x m-k+1 (h)} ...

[0138] X m ={x m (hn),x m (h-n+1),...,x m (h)}

[0139] Where Θ is the eigenvalue sample set; X1 is the first subset formed based on the temperature time series of the target station; X2~X m-k X is the 2nd to mkth subsets formed based on the auxiliary variable time series of the target station; m-k+1 ~X m The m-k+1 to m-th subsets are formed based on the temperature time series of similar stations, and m is the number of variable dimensions of the eigenvalue sample;

[0140] x1(hn), x1(h-n+1), ..., x1(h) are the eigenvalue samples of the temperature of the target station corresponding to time hn, h-n+1, ​​..., h, respectively, and n is the number of lag time dimensions of the eigenvalue samples; x2(hn), x2(h-n+1), ..., x2(h) are the eigenvalue samples of the first auxiliary variable of the target station corresponding to time hn, h-n+1, ​​..., h, respectively;

[0141] x m-k (hn),x m-1 (h-n+1),...,x m-1 (h) are the characteristic value samples of the m-2th auxiliary variable of the target station corresponding to time hn, h-n+1, ​​..., h, and the total number of auxiliary variables is m-2;

[0142] x m-k+1 (hn),x m-k+1 (h-n+1),...,x m-k+1(h) are the characteristic value samples of the temperature of the first similar station corresponding to time hn, h-n+1,…,h respectively;

[0143] x m (hn),x m (h-n+1),...,x m (h) are the characteristic value samples of the temperature at the k-1th similar station corresponding to time hn, h-n+1, ​​..., h, and the total number of similar stations is k-1;

[0144] Then the relationship between the eigenvalue sample and the actual temperature is:

[0145]

[0146] Where y(h+k) is the actual temperature value at the prediction time h+k; f(*) is the relationship model between the eigenvalue sample and the actual temperature; ε is the error.

[0147] Example 6:

[0148] This embodiment is further designed based on the first embodiment in that the temperature prediction model in this embodiment adopts a nonlinear convex programming integrated deep basic model with constraints;

[0149] The objective function of the temperature prediction model is:

[0150]

[0151] Where Z is the difference between the predicted temperature and the actual temperature; α is the α-norm; w j is the integration weight of the deep base model j to the minimization function, j∈[1,k], k is the total number of deep base models; y i is the actual temperature value of the predicted value i; is the temperature prediction value of deep base model j for the predicted value i, n is the total number of values ​​to be predicted; λ is the regularization parameter used to balance the two objective terms of the minimization function; w j The norm of

[0152] The constraints of the temperature prediction model are:

[0153]

[0154] The loss function of the basic model of the temperature prediction model is:

[0155]

[0156]

[0157] Where ρ(r) is the loss function value corresponding to the model error r; r is the difference between the temperature prediction value of the basic model and the actual temperature value, that is, the basic model error; c is a constant, also called the robust scaling constant.

[0158] Embodiment seven:

[0159] This embodiment is further designed on the basis of Example 6 in that the deep basic model integrated in the temperature prediction model in this example includes a gated recursive unit (GRU), a bidirectional gated recursive unit (BiGRU), a convolutional neural network (CNN), a long short-term memory network (LSTM), a convolutional neural network coupled with a long short-term memory network CNN-LSTM, a convolutional neural network coupled with a bidirectional long short-term memory network CNN-BiLSTM, a convolutional neural network coupled with a gated recursive unit CNN-GRU, and a convolutional neural network coupled with a bidirectional gated recursive unit CNN-BiGRU.

[0160] Embodiment 8:

[0161] This embodiment is further designed based on the sixth embodiment in that the temperature prediction model in this embodiment uses the standard deviation index SD (standard deviation, abbreviated as SD), the root mean square error index RMSE (Root Mean Square Error, abbreviated as RMSE) and the correlation coefficient index CC (correlation coefficient, abbreviated as CC) to evaluate the model prediction accuracy. The expressions of the standard deviation index SD, the root mean square error index RMSE and the correlation coefficient index CC are as follows:

[0162]

[0163]

[0164]

[0165] Where x i is the difference between the predicted temperature value and the actual temperature value of the predicted value i; i is the actual temperature value of the predicted value i; is the mean of the actual temperature value; P i is the predicted value of the temperature of the predicted value i; μ is the mean of the difference between the predicted value of the temperature of the predicted value i and the actual value of the temperature, n is the number of values ​​to be predicted.

[0166] Embodiment 9:

[0167] This embodiment is further designed based on the first embodiment in that a process interpretable analysis is performed on the initial regional temperature forecast results to obtain a specific method for the process interpretable contribution rate of different eigenvalue samples for measuring the importance of the eigenvalue samples to the temperature forecast:

[0168] Assume x∈R M×N is a time series with M variables and N time steps (lag time) (e.g., temperature). Based on the relationship between meteorological characteristics and lag time in the eigenvalue sample, a set of variable and time step combinations Δ is formed.

[0169] Δ={(i,t):1≤i≤M,1≤t≤N}

[0170] Where (i, t) is the combination of the i-th variable and the t-th time step; M is the total number of variables, corresponding to the number of variable dimensions in the eigenvalue sample; N is the total number of time steps, corresponding to the number of lag time dimensions in the eigenvalue sample;

[0171] Calculate the ProcessSHAP value for different combinations of variables and time steps. The ProcessSHAP value P for the combination of the i-th variable and the t-th time step is (i,t) ,

[0172]

[0173]

[0174] Where, is the characteristic function, which means that when the input X * The predicted output when there are only variable time pairs in the set S; Δ i is the set of all time steps of variable i; set S is a subset of all features; x S is a subvector of x, representing the features in the set S; f(·) represents the prediction model; x represents the input of the prediction model;

[0175] The sum of ProcessSHAP values ​​f(x * ), Where, P 0,0 This is the output of the feature function when all features do not exist. The sum of the feature function outputs for all combined feature sets;

[0176] For the original / classic CNN convolutional layer, assume that the activation function of the lth convolutional layer is θ (l) , by changing θ(l) With the convolution kernel w (l) Convolution gets the output O (l) The receptive field size of the convolution kernel is A×B, and the sliding step is z. The bias term is expressed as The classic or original convolutional layer is calculated as follows:

[0177]

[0178] Here, x represents the x-th row in the data matrix, y represents the y-th column in the data matrix, and the operator * represents the convolution operation.

[0179] Normalize the ProcessSHAP values ​​under different combinations of variables and time steps to obtain the process explanation contribution rate of different combinations of variables and time steps. For the combination of the i-th variable and the t-th time step, the process explanation contribution rate C (i,t) The calculation formula is as follows:

[0180]

[0181] The process based on the combination of variables and time steps can explain the contribution rate, and the process of obtaining different eigenvalue samples to measure the importance of eigenvalue samples to temperature prediction can explain the contribution rate.

[0182] Embodiment 10:

[0183] This embodiment is further designed based on the first embodiment in that, in this embodiment, based on the process explainable contribution rate, the specific steps of updating the importance of salient feature samples to temperature prediction in the regional temperature prediction model that has been initially trained include:

[0184] The process explainable contribution rate is coupled as a weight to the convolution kernel of the convolution layer in the regional temperature prediction model that has been initially trained, thereby updating the importance of the salient feature samples to the temperature prediction. The coupled convolution layer is calculated as follows:

[0185]

[0186] Where x is the xth row in the data matrix, y is the yth column in the data matrix, the operator * represents the convolution operation; l is the number of convolution layers; θ (l) is the activation function of the lth convolutional layer; w (l) is the convolution kernel of the lth convolution layer; A×B is the receptive field size of the convolution kernel, and z is the sliding step size; is the bias term of the lth convolutional layer; C (i,t) is the process-interpretable contribution of the eigenvalue sample corresponding to the combination of the i-th variable and the t-th time step.

[0187] Example 11:

[0188] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to execute the method described in any of the above embodiments.

[0189] A computer-readable storage medium stores a computer program, which implements the steps of the method described in any of the above embodiments when executed by a processor.

[0190] Application examples:

[0191] This example uses the temperature data from the national automatic station in Anhui Province, China to evaluate the prediction accuracy of the prediction method of the present invention.

[0192] 1) This example introduces the study area and stations. Figure 3(a) shows the distribution of the 81 national automatic meteorological stations in Anhui Province used in this study. The color scale represents the station's altitude (unit: meters). Meteorological data includes temperature, air pressure, relative humidity, vapor pressure, wet-bulb temperature, and wind speed. This example uses data from four representative stations in Anhui Province (Figure 3(b)) from 00:00 on January 1, 2007, to 23:00 on December 31, 2016, to develop a temperature prediction model. Station number 58118 (Mengcheng, Anhui) has the following geographic information: longitude 116.53, latitude 33.278. Station number 58321 (Hefei, Anhui) has the following geographic information: longitude 117.3, latitude 31.78. Station number 58112 (Tianzhu Mountain, Anhui) has the following geographic information: longitude 116.46, latitude 30.733. Station ID 58423 (Jiuhua Mountain, Anhui Province) Geographic information: Longitude 117.78, Latitude 30.483. The color scale indicates altitude (in meters).

[0193] 2) The effectiveness of the loss function of the model constructed in this invention was analyzed. In this example, the temperature prediction results were obtained using the mean absolute error (MAE), mean square error (MSE), and the modified Huber function in M-estimation (one of the loss functions used in this invention) as the loss function of the basic model. The accuracy of the model was evaluated using the root mean square error (in degrees Celsius) to predict the temperature in the next 1, 2, 3, 4, 5, 6, 8, 10, 12, 15, 18, and 24 hours with an 8-hour lag time. Figure 4 The results of temperature prediction for different time periods at station 58321 based on different loss functions are given. Figure 4 It can be seen that the overall temperature prediction accuracy using the loss function of the present invention is higher than that of "MAE" and "MSE". The reason why the temperature prediction accuracy is close in some areas may be attributed to the selection of the robust scaling constant in the loss function.

[0194] 3) Analysis of the effectiveness of the prediction method of the present invention. The proposed model was applied to temperature prediction at a typical automatic station in Anhui Province, China, on an hourly scale. The accuracy of temperature prediction using baseline models (LSTM, BiLSTM), basic models (GRU, BiGRU, CNN-LSTM, CNN-BiLSTM, CNN-GRU, CNN-BiGRU), and deep ensemble learning was compared. Figure 5(a) to Figure 5(d) The station numbers 58112, 58118, 58321, and 58423 are given respectively, and the standard deviation values ​​are used to measure the temperature accuracy of different models for different time periods. Figure 5(a) to Figure 5(d) It can be seen that compared with the temperature prediction accuracy of the baseline model and the basic model, the temperature prediction accuracy of the multi-model deep integrated learning in the present invention is the highest overall. Figure 6(a) to Figure 6(d) The scatter plots of the predicted and observed temperature values ​​for Hefei (station number 58321) are given. This example only shows the scatter plots of the predicted and observed temperature values ​​for the test set 1, 3, 6, and 12 hours. The color scale indicates the number of sample points. Figure 6(a) to Figure 6(d) As can be seen, the correlation coefficients between the 1-, 3-, 6-, and 12-hour forecasts and the observed values ​​are all over 0.97. These scatter plots also show a strong correlation between the observed and predicted values ​​of temperature.

[0195] 4) Analyze the process interpretability analysis steps in the present invention. In order to clearly reveal the influence or importance of different features on the temperature prediction results, this example selects the process interpretability analysis results of the SHAP deep interpreter with a lag time of 8 hours to predict the next 2 hours. Further analysis shows that the results of predicting 1 hour and 3 hours with a lag time of 8 hours are similar (Figure omitted). Different from the commonly used SHAP tree interpreter (Tree Explainer), Figure 7 The results of SHAP Deep Explainer are given to reveal the model mechanism of the deep model in predicting temperature. The test set samples are selected for global interpretation and the feature density scatter plots are drawn. Figure 7 . Figure 7 The horizontal axis is the SHAP value, and each row represents a feature.

[0196] Among them, TST represents the historical temperature value of the station to be predicted. Ri represents the temperature of the reference station i (i = 1, 2, ... 5). P represents air pressure. RH represents relative humidity. WVP represents water vapor pressure. WBT represents wet bulb temperature. WS represents wind speed. H represents time. t-1 represents a lag of 1 hour. t-2 represents a lag of 2 hours. The description of related variables is similar. "SHAP value (impact on model output)" represents SHAP value (impact on model output). "Feature value" represents feature value. Low represents low value. High represents high value. "SHAP DeepExplainer" represents SHAP deep explainer. The color table represents the size of the value. Figure 7 As can be seen from the DeepExplainer, the first feature "TST(t-1)" for station 58112 represents the temperature value one time step prior to the target station's predicted temperature time point. The feature "H(t-8)" represents the "time variable" for the eight time steps prior to the target station's predicted temperature time point. This achieves process-based interpretability in the temperature model. Similar analysis can be performed for other stations and variables. Figure 7 This reveals that the deep model relies heavily on auxiliary information other than the target station's temperature. The "reference station" and "time" features are also important.

[0197] 5) The short-term temperature intelligent prediction method of the present invention is applied to precipitation in some parts of China caused by low-level shear lines. From March 7 to March 10, 2016, due to the influence of the low-level shear line formed by the intersection of the fast-moving southward cold air and the strong southerly warm and humid air flow, a severe convective weather process was triggered in some parts of China, resulting in short-term heavy rainfall and thunderstorms. From March 7 to 8, under the influence of the shear line, thunderstorms with wind speeds of 18-24m / s occurred in the southwest of Anhui Province. From March 8 to 9, thunderstorms with wind speeds of 18-24m / s occurred in the southwest of Anhui Province, southern Jiangsu Province and other areas. Specifically, from 20:00 on March 7 to 14:00 on March 10, short-term heavy rainfall occurred in southern Anhui Province and other places; from the afternoon to the first half of the night on the 8th, thunderstorms and winds occurred in southern Anhui Province, Shanghai and other places, with the maximum wind speed reaching level 9 (24m / s). Figure 8(a) to Figure 8(d) Time series plots of the predicted and observed temperatures for 1 hour, 2 hours, and 3 hours ahead at Tianzhushan Station (station number 58112), Mengcheng Station (station number 58118), Hefei Station (station number 58321), and Jiuhuashan Station (station number 58423) in Anhui Province during this period are presented. Time series plots of the 1-hour precipitation observed at the corresponding stations are also presented.

[0198] Table 1 Analysis of temperature prediction accuracy at different stations and different prediction times (correlation coefficient)

[0199]

[0200] Depend on Figure 8(a) to Figure 8(d) As shown in Table 1, the predicted and observed temperature values ​​at different stations are highly consistent. For the stations selected for this study, the temperature correlation coefficients all exceeded 0.97.

[0201] 6) Further analysis of the process heatmap of the DeepExplainer model results for predicting the next 2 hours with a time lag of 8 hours. Figure 9 As shown in the figure, this section analyzes the process-based interpretation of temperature forecasts. TST represents the historical temperature value at the station being predicted. Ri represents the temperature at reference station i (i = 1, 2, ..., 5). P represents air pressure. RH represents relative humidity. WVP represents water vapor pressure. WBT represents wet-bulb temperature. WS represents wind speed. H represents time. "Sum of 87 other features" represents other feature values ​​not shown in the figure. Instances represents instances. "SHAP value" represents the SHAP value. t-1 represents a lag of 1 hour. t-2 represents a lag of 2 hours. The explanations for related variables are similar. The color scale indicates the magnitude of the value.

[0202] Depend on Figure 9 As can be seen, the key characteristic variables in the temperature forecast process for different stations are relatively consistent with the results of the characteristic density scatter plot. In the "process-based" analysis, characteristic variables closer to the prediction time have a greater impact on the final prediction results. This analysis provides the basis for the secondary training of the proposed method (incorporating the convolutional layer calculation of the characteristic contribution rate for process interpretability) and also lays the foundation for the subsequent addition of "target adaptive observation variables" to the temperature forecast process.

[0203] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for intelligent prediction of regional short-term temperature based on process-interpretable analysis, wherein the region is provided with a target station and a reference station, characterized in that: include: Obtaining the temperature time series and auxiliary variable time series of the target station and the temperature time series of the reference station; Based on the similarity between the temperature time series of the target station and the reference station, similar stations are screened among the reference stations while considering the nonlinear relationship between the temperature time series and the temperature time series; Based on the temperature time series of the selected similar stations and the temperature time series and auxiliary variable time series of the target station, the characteristic value samples are constructed with the set lag time; Based on the characteristic value samples, a pre-built regional temperature prediction model is initially trained, and an initial regional temperature prediction result is obtained based on the regional temperature prediction model that has been initially trained; Performing a process interpretable analysis on the initial regional temperature forecast results to obtain process interpretable contribution rates of different eigenvalue samples for measuring the importance of the eigenvalue samples to the temperature forecast; Based on the explained contribution rate of the process, the regional temperature prediction model that has been initially trained is updated with respect to the importance of the salient feature samples to the temperature prediction; Based on the characteristic value samples, the updated regional temperature prediction model is trained a second time, and based on the regional temperature prediction model after the second training, a second regional short-term temperature prediction result is obtained and used as the final regional short-term temperature intelligent prediction result; The specific method of performing process interpretable analysis on the initial regional temperature forecast result to obtain process interpretable contribution rates of different eigenvalue samples for measuring the importance of eigenvalue samples to temperature forecast is as follows: Based on the relationship between the meteorological characteristics and the lag time in the characteristic value samples, a set Δ of variable and time step combinations is formed, Δ={(i,t):1≤i≤M,1≤t≤N} Where (i, t) is the combination of the i-th variable and the t-th time step; M is the total number of variables, corresponding to the number of variable dimensions in the eigenvalue sample; N is the total number of time steps, corresponding to the number of lag time dimensions in the eigenvalue sample; Calculate the ProcessSHAP value for different combinations of variables and time steps. The ProcessSHAP value P for the combination of the i-th variable and the t-th time step is (i,t) , Where, is the characteristic function, which means that when the input X * The predicted output when there are only variable time pairs in the set S; Δ i is the set of all time steps of variable i; set S is a subset of all features; x S is a subvector of x, representing the features in the set S; f(·) represents the prediction model; x represents the input of the prediction model; The sum of ProcessSHAP values ​​f(x * ), Where, P 0,0 The output of the feature function when all features do not exist; The sum of the feature function outputs for all combined feature sets; Normalize the ProcessSHAP values ​​under different combinations of variables and time steps to obtain the process explanation contribution rate of different combinations of variables and time steps. For the combination of the i-th variable and the t-th time step, the process explanation contribution rate C (i,t) The calculation formula is as follows: The process based on the combination of variables and time steps can explain the contribution rate, and the process of obtaining different eigenvalue samples to measure the importance of eigenvalue samples to temperature prediction can explain the contribution rate.

2. The regional short-term temperature intelligent prediction method based on process interpretable analysis according to claim 1 is characterized in that: The auxiliary variables include air pressure, relative humidity, water vapor pressure, wet bulb temperature, wind speed and diurnal variation time of air temperature.

3. The method for intelligent prediction of regional short-term temperature based on process-interpretable analysis according to claim 1 is characterized in that: The specific steps of obtaining the temperature time series and auxiliary variable time series of the target station and the temperature time series of the reference station include: Obtaining historical meteorological characteristic values ​​of the target station and historical temperature values ​​of the reference station, wherein the meteorological characteristics include temperature and auxiliary variables; The historical meteorological characteristic values ​​of the target station and the historical temperature values ​​of the reference station are respectively subjected to preprocessing including missing value interpolation processing and feature normalization processing, and sorting processing by timestamp to form the temperature time series and auxiliary variable time series of the target station and the temperature time series of the reference station; wherein, The specific method of the missing value interpolation process is: using the K-nearest neighbor algorithm for interpolation; The specific method of the feature normalization processing is: using the Max-Min normalization method to perform feature normalization processing on historical meteorological feature values ​​and historical temperature values. The specific processing formula is as follows: Where y is the historical meteorological characteristic value or historical temperature value after feature normalization; x is the original value of the historical meteorological characteristic value or historical temperature value; x min is the minimum value among the original values ​​of historical meteorological characteristic values ​​or historical temperature values; x max It is the maximum value among the original values ​​of historical meteorological characteristic values ​​or historical temperature values.

4. The regional short-term temperature intelligent forecasting method based on process interpretable analysis according to claim 1 is characterized in that: The specific method of screening similar stations in the reference station based on the similarity between the temperature time series of the target station and the reference station while considering the nonlinear relationship between the temperature time series and the temperature time series includes: Calculate the similarity between the temperature time series of the target station and the reference station, and select several reference stations with the highest similarity to the target station as similar stations; the similarity calculation formula is as follows: where p(x1) is the marginal distribution of the temperature time series x1 of the target station; p(y1) is the marginal distribution of the temperature time series y1 of the reference station; p(x1,y1) is the joint distribution of the temperature time series x1 of the target station and the temperature time series y1 of the reference station; I(x1;y1) is the similarity between the temperature time series x1 of the target station and the temperature time series y1 of the reference station, that is, the relative entropy of the joint distribution and marginal distribution between the temperature time series x1 of the target station and the temperature time series y1 of the reference station.

5. The method for intelligent prediction of regional short-term temperature based on process interpretable analysis according to claim 1 is characterized in that: The specific method of constructing a characteristic value sample with a set lag time based on the screened temperature time series of similar stations, the temperature time series of the target station, and the auxiliary variable time series includes: Set the lag time n that takes into account both prediction accuracy and task requirements; Based on the temperature time series of the selected similar stations, the temperature time series of the target station, and the auxiliary variable time series, a sample with characteristic values ​​is constructed: Θ={X1,X2,...,X m-1 ,X m } X1={x1(hn),x1(h-n+1),...,x1(h)} X2={x2(hn),x2(h-n+1),...,x2(h)} ... X m-k ={x m-k (h-n),x m-k (h-n+1),...,x m-k (h)} X m-k+1 ={x m-k+1 (h-n),x m-k+1 (h-n+1),...,x m-k+1 (h)} ... X m ={x m (h-n),x m (h-n+1),...,x m (h)} Where Θ is the eigenvalue sample set; X1 is the first subset formed based on the temperature time series of the target station; X2~X m-k X is the 2nd to mkth subsets formed based on the auxiliary variable time series of the target station; m-k+1 ~X m The m-k+1 to m-th subsets are formed based on the temperature time series of similar stations, and m is the number of variable dimensions of the eigenvalue sample; x1(hn), x1(h-n+1), ..., x1(h) are the eigenvalue samples of the temperature of the target station corresponding to time hn, h-n+1, ​​..., h, respectively, and n is the number of lag time dimensions of the eigenvalue samples; x2(hn), x2(h-n+1), ..., x2(h) are the eigenvalue samples of the first auxiliary variable of the target station corresponding to time hn, h-n+1, ​​..., h, respectively; x m-k (hn),x m-k (h-n+1),...,x m-k (h) are the characteristic value samples of the mk-1th auxiliary variable of the target station corresponding to time hn, h-n+1, ​​..., h, and the total number of auxiliary variables is mk-1; x m-k+1 (hn),x m-k+1 (h-n+1),...,x m-k+1 (h) are the characteristic value samples of the temperature of the first similar station corresponding to time hn, h-n+1,…,h respectively; x m (hn),x m (h-n+1),...,x m (h) are the characteristic value samples of the temperature at the k-th similar station corresponding to time hn, h-n+1, ​​..., h, and the total number of similar stations is k; Then the relationship between the eigenvalue sample and the actual temperature is: Where y(h+k) is the actual temperature value at the prediction time h+k; f(*) is the relationship model between the eigenvalue sample and the actual temperature; ε is the error.

6. The method for intelligent prediction of regional short-term temperature based on process-interpretable analysis according to claim 1 is characterized in that: The temperature prediction model adopts a nonlinear convex programming with constraints and an integrated deep foundation model; The objective function of the temperature prediction model is: Where Z is the difference between the predicted temperature and the actual temperature; α is the α-norm; w j is the integration weight of the deep base model j to the minimization function, j∈[1,k], k is the total number of deep base models; y i is the actual temperature value of the predicted value i; is the temperature prediction value of deep basic model j for the predicted value i, and n is the total number of values ​​to be predicted; λ is the regularization parameter, which is used to balance the two objective terms of the minimization function; w j The norm of The constraints of the temperature prediction model are: The loss function of the basic model in the temperature prediction model is: Where ρ(r) is the loss function value corresponding to the basic model error r; r is the difference between the temperature prediction value of the basic model and the actual temperature value, that is, the basic model error; c is the robust scaling constant.

7. The method for intelligent prediction of regional short-term temperature based on process-interpretable analysis according to claim 6 is characterized in that: The deep basic models integrated in the temperature prediction model include gated recurrent units, bidirectional gated recurrent unit convolutional neural networks, long short-term memory networks, convolutional neural networks coupled with long short-term memory networks, convolutional neural networks coupled with bidirectional long short-term memory networks, convolutional neural networks coupled with gated recurrent units, and convolutional neural networks coupled with bidirectional gated recurrent units.

8. The method for intelligent prediction of regional short-term temperature based on process-interpretable analysis according to claim 6 is characterized in that: The temperature prediction model uses the standard deviation index SD, the root mean square error index RMSE and the correlation coefficient index CC to evaluate the model prediction accuracy. The expressions of the standard deviation index SD, the root mean square error index RMSE and the correlation coefficient index CC are as follows: Where x i is the difference between the predicted temperature value and the actual temperature value of the predicted value i; i is the actual temperature value of the predicted value i; is the mean of the actual temperature value; P i is the predicted temperature value of the predicted value i; μ is the mean of the difference between the predicted temperature value of the predicted value i and the actual temperature value, n is the number of values ​​to be predicted.

9. The method for intelligent prediction of regional short-term temperature based on process-interpretable analysis according to claim 1 is characterized in that: The specific steps of updating the importance of salient feature samples to temperature prediction in the regional temperature prediction model that has been initially trained based on the explainable contribution rate of the process include: The process explainable contribution rate is coupled as a weight to the convolution kernel of the first convolution layer in the regional temperature prediction model that has been initially trained, thereby updating the importance of the salient feature samples to the temperature prediction. The coupled convolution layer is calculated as follows: Where x is the xth row in the data matrix, y is the yth column in the data matrix, the operator * represents the convolution operation; l is the number of convolution layers; θ (l) is the activation function of the lth convolutional layer; w (l) is the convolution kernel of the lth convolution layer; A×B is the receptive field size of the convolution kernel, and z is the sliding step size; is the bias term of the lth convolutional layer; C (i,t) is the process-interpretable contribution of the eigenvalue sample corresponding to the combination of the i-th variable and the t-th time step.

10. The method for intelligent prediction of regional short-term temperature based on process-interpretable analysis according to claim 1 is characterized in that: In the secondary training, the model parameters of all convolutional layers are fixed, and only the model parameters of one fully connected layer are adjusted.

Images