CN115526413A - A Forecasting Method of Daily Maximum Air Temperature Based on Fully Connected Neural Network - Google Patents

Abstract

The invention relates to a forecasting method based on the daily maximum temperature of a fully-connected neural network, which comprises the following steps: s1, preprocessing observation data and numerical forecast data; s2, dividing the preprocessed data into a training set, a verification set and a test set, and respectively carrying out normalization processing on the characteristics and target execution data on the training set, the verification set and the test set; s3, constructing a full-connection neural network model combined with the embedded layer, setting hyper-parameters of the model and training the model; and S4, evaluating the model after training is finished, and predicting the highest air temperature through the model. According to the invention, a time lag variable is selected, meanwhile, in order to identify time information of different sites and samples in a data set, an auxiliary variable is designed, and an embedded layer is designed aiming at two classifications of site numbers and months, so that seasonal and regional modeling is avoided.

Description

A Forecasting Method of Daily Maximum Temperature Based on Fully Connected Neural Network

technical field

The invention relates to the technical field of meteorological monitoring, in particular to a method for forecasting daily maximum temperature based on a fully connected neural network.

Background technique

Objective forecasting of maximum temperature is an important content of weather forecasting, which has important practical application significance, such as traffic control, weather forecasting, agricultural production and environmental monitoring, etc. Usually, the forecasting method of forecasting the daily maximum temperature generally adopts the forecasting method of the atmospheric numerical model. For example, the Integrated Forecast System model of the European Center for Medium-Range Weather Forecasting (referred to as the IFS model, the same below), the regional numerical forecast model of the China Meteorological Administration, and so on. Affected by complex factors such as atmospheric dynamic processes, physical processes, and local terrain and landforms, there are often deviations in the forecast of near-surface elements (including maximum air temperature) forecasted by atmospheric numerical models, especially when turning weather occurs. The error between the actual observations is larger, so it is still a challenge to achieve the refinement and accuracy of the daily maximum temperature forecast. Although using traditional statistical methods to correct the forecast error of the numerical model can improve the forecast effect to a certain extent, there are limitations, such as modeling for a single spatial point, which not only requires a large amount of calculation, but also cannot consider the interaction of meteorological elements between meteorological stations. influences. Therefore, how to solve the existing problems in numerical model forecasting using traditional statistical methods needs to be considered at this stage.

It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to those of ordinary skill in the art.

Contents of the invention

The purpose of the present invention is to overcome the shortcoming of prior art, provide a kind of forecasting method based on fully connected neural network daily maximum temperature, solve the existing method of utilizing atmospheric numerical model to forecast daily maximum temperature to be affected by atmospheric dynamic process, physical process and The complex influence of local topography and geomorphology leads to the problem that the forecast of near-surface elements of the atmospheric numerical model, including the maximum temperature, often has deviations.

The object of the present invention is achieved through the following technical solutions: a method for forecasting the daily maximum temperature based on a fully connected neural network, the forecast method comprising:

Step 1, preprocessing the observation data and numerical forecast data;

Step 2. Divide the preprocessed data into training set, verification set and test set, and perform normalization processing on the features and target execution data on the training set, verification set and test set respectively;

Step 3. Construct a fully connected neural network model combined with an embedding layer, and set the hyperparameters of the model and train the model;

Step 4. Evaluate the model after training, and predict the maximum temperature through the model.

The preprocessing of the observation data and numerical forecast data includes the following:

S11. Perform data cleaning on the observation data and generate the target file: eliminate the wrong or missing sites, select the maximum temperature observation data of the site with a complete time series, and create the target data set;

S12. Read and convert the data format of the historical numerical model: read the GRIB format data of the numerical forecast model, unify the horizontal resolution of the ground layer forecast data and the upper-altitude layer forecast data, and linearly interpolate a small number of missing forecast values in the time dimension The method performs interpolation, and outputs the historical numerical forecast data in nc format;

计算涡度平流，通过计算公式

计算温度平流，其中，v代表风速，

是等压面上的拉普拉斯算子，ζ是相对涡度，f是行星涡度，T为温度；S13. Calculate the upper-altitude combination data of the numerical model: based on the upper-altitude layer data of the numerical model, through the calculation formula

Calculate the eddy advection, through the calculation formula

Calculate the temperature advection, where v represents the wind speed,

is the Laplace operator on the isobaric surface, ζ is the relative vorticity, f is the planetary vorticity, T is the temperature;

S14, interpolating the data on the numerical model grid point to the weather station site: adopting the bilinear interpolation method to interpolate the ground data, high-altitude data and high-altitude combined data on the numerical model grid point to the ground weather station site, and interpolating The final numerical model data and target data are output in one file;

S15. Build model input features: According to the forecast experience, select the predictors related to the daily maximum temperature forecast on the ground layer and the upper layer, and select the predictors related to the forecast quantity as the model features through the correlation analysis method and the mutual information value method .

The selection of the predictors related to the predictor as model features through the correlation analysis method and the mutual information value method includes:

选择出与目标相关系数大于0.3，且达到0.05显著性水平的因子，其中，X是观测值，代表目标，Y是预报值，代表特征，cov(X,Y)为X,Y的协方差，σ_X是X的标准差，μ_X是X的期望E(X)；Correlation analysis method: through the correlation coefficient formula

Select a factor with a correlation coefficient greater than 0.3 and reach a significance level of 0.05, where X is the observed value representing the target, Y is the predicted value representing the feature, and cov(X,Y) is the covariance of X and Y, σ _X is the standard deviation of X, μ _X is the expected E(X) of X;

Mutual information value method: By calculating the mutual information value between the target and the feature, the nonlinear correlation between the feature and the target is measured, and the joint distribution of two random variables (X, Y) is set to p(x, y), and the marginal distribution They are respectively p(x), p(y), mutual information I(X; Y) is the relative entropy of the joint distribution p(x, y) and the marginal distribution p(x), p(y), and the mutual information value is expressed as

The time lag variable is selected as the maximum temperature observation data with a lag of N days and N+1 days, and an auxiliary variable is selected to identify the time information of different stations and samples in the data set to complete the creation of the feature data set.

Described auxiliary variable comprises season, month, weather station number, station longitude, station latitude and station altitude; Number, season and month are used as categorical variables.

The constructed fully connected neural network model combined with embedding layers includes an input layer, an embedding layer, a connection layer, a hidden layer and an output layer; the input layer includes label encoding and one-hot encoding, and the station number and month of the weather station in the classification variable Input into the label encoding, the season in the categorical variable is input into the one-hot encoding; the embedding layer is used to process the categorical variable passed by the label encoding in the input layer; the connection layer is used to connect the features processed by the embedding layer, the independent Hot-encoded processed features and other features; the hidden layer consists of multiple neurons for feature extraction and learning; finally output through the output layer.

The hyperparameters of the model are set and the model is trained including:

Set the embedding vector dimension in the embedding layer to 8, the number of hidden layers to 1, the number of neurons in the hidden layer to 64, the activation function to ReLU, the number of neurons in the output layer to 1, and the activation function is Linear;

Set the loss function to mean absolute error, use this regularization method to prevent the model from overfitting, set the number of model iterations to 50 times, if the verification error does not decrease but increases for 3 consecutive times during training, stop training, and the model training is complete , then apply the model to the validation set, calculate the model error on the validation set, perform hyperparameter selection by comparing the validation set to obtain the best model, and save the model with the best performance on the validation set.

The evaluation of the model at the end of the training and the prediction of the maximum temperature through the model include:

Load the model that has been saved after training, input the characteristic data in the test set in step S2 into the model to output the predicted value of daily maximum temperature, and compare the predicted value output by the model with the actual observed value to make the model independent of the forecast outside the training set performance evaluation;

When there is a new data sample to predict the daily maximum temperature, the new data sample is preprocessed according to step S1, and then the processed data is input into the model to obtain the predicted value of the maximum temperature.

The present invention has the following advantages: a method for forecasting the daily maximum temperature based on a fully connected neural network, which can identify the time and space information of samples and process multi-category variables, and the fully connected neural network model with an embedded layer can realize more accurate daily temperature prediction. Forecasting the maximum temperature; the time lag variable is selected, and auxiliary variables are designed to identify the time information of different stations and samples in the data set, and an embedding layer is designed for the two classifications of station number and month to avoid seasons and regional modeling.

Description of drawings

Fig. 1 is a schematic flow sheet of the present invention;

Fig. 2 is a schematic diagram of the principle of bilinear interpolation;

Fig. 3 is a structural diagram of a fully connected neural network model of the present invention;

Fig. 4 is the schematic diagram that the present invention method forecasts maximum air temperature;

Figure 5 is a schematic diagram of the daily maximum temperature forecast by the existing numerical forecast IFS model method.

detailed description

In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only It is a part of the embodiments of this application, not all of them. The components of the embodiments of the application generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the application provided in conjunction with the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative efforts belong to the scope of protection of the present application. The present invention will be further described below in conjunction with the accompanying drawings.

As shown in Figure 1, the present invention relates to a method for forecasting the daily maximum temperature based on a fully connected neural network, which can identify the spatiotemporal information of samples and process multi-category variables. Through a fully connected neural network model with an embedded layer, it aims at To solve the technical problem of forecasting the daily maximum temperature of meteorological stations in China, which is more accurate than the daily maximum temperature forecast of the integrated forecasting system of the European Center for Medium-Range Weather Forecasting; specifically, the following steps are included:

S1. Preprocessing the observation data and numerical forecast data;

S11. Perform data cleaning on the observation data and generate target files:

Eliminate erroneous or missing meteorological station sites, select the maximum temperature observation data with a complete time series, and create a target data set (number of dates * number of stations, 1).

S12. Read and convert historical numerical mode data format:

Read the data in the GRIB format of the numerical prediction model, and unify the horizontal resolution of the ground-level forecast data and the upper-air layer forecast data. For a small number of missing forecast values in the time dimension, the linear interpolation method is used to interpolate, and the historical numerical forecast data is output in nc format.

S13. Calculate the high-altitude combination data of the numerical model:

其中，v代表风速，

是等压面上的拉普拉斯算子，ζ是相对涡度，f是行星涡度(科里奥利参数)。温度平流公式为：

其中，v代表风速，T为温度。Based on the upper-air layer data of the numerical model, the eddy advection and temperature advection are calculated. The calculation formula is:

Among them, v represents the wind speed,

is the Laplace operator on the isobaric surface, ζ is the relative vorticity, and f is the planetary vorticity (Coriolis parameter). The temperature advection formula is:

Among them, v represents the wind speed, and T is the temperature.

S14, interpolating the data on the grid points of the numerical model to the weather station site:

Using the bilinear interpolation method, the ground data, upper-air data and upper-air combination data on the numerical model grid points are interpolated to the ground national weather station site, and the interpolated numerical model data and target data are output as a file. Among them, the principle of bilinear interpolation is as follows:

As shown in Figure 2, for example, the target to be obtained is the value of the unknown function f at point P=(x,y), and the known function f is at Q ₁₁ =(x ₁ ,y ₁ ), Q ₁₂ =(x ₁ ,y ₂ ), Q ₂₁ =(x ₂ ,y ₁ ) and Q ₂₂ =(x ₂ ,y ₂ ) four points.

First perform linear interpolation in the X direction to obtain R ₁ and R ₂ , the specific calculation formula is as follows:

Wherein, R ₁ =(x,y ₁ )R ₂ =(x,y ₂ ). Then linear interpolation in the y direction to calculate the value of point P, the calculation formula is as follows:

S15. Build model input features:

According to the forecast experience, the predictors related to the daily maximum air temperature forecast are selected for the ground layer and the upper layer. Using the correlation analysis method and the mutual information value method, the predictors related to the forecast quantity are selected as the model features; the correlation analysis method and the mutual information value method are explained as follows:

(1) Correlation analysis method: the features that have a certain linear relationship with the target can be selected. The method selects the factors whose correlation coefficient with the target is greater than 0.3 and reaches the significance level of 0.05. Among them, the correlation coefficient formula is:

X is the observed value (target), Y is the predicted value (feature), cov(X,Y) is the covariance of X and Y, σ _X is the standard deviation of X, and μ _X is the expected E(X) of X.

(2) Mutual information value method: By calculating the mutual information value between the target and the feature, the non-linear correlation between the feature and the target is measured. Let the joint distribution of two random variables (X, Y) be p(x, y), the marginal distributions are respectively p(x), p(y), and the mutual information I(X; Y) is the joint distribution p(x, y) and the relative entropy of the marginal distributions p(x), p(y). The mutual information value is expressed as I(X; Y), and its formula is as follows:

In order to avoid the model from relying on the numerical model forecast, the time lag variable selected in the present invention is the observed data of maximum temperature with a lag of 1 day and a lag of 2 days. At the same time, in order to identify the time information of different stations and samples in the data set, auxiliary variables are designed. The selected auxiliary variables include season, month, weather station number, station longitude, station latitude, and station altitude. The present invention regards the station longitude, station latitude and station altitude as numerical variables, and station number, season and month as classification variables. Finish creating feature data set (number of dates * number of sites, number of features).

One-hot encoding is used for seasonal variables. For station number and month, first use label encoding for numerical mapping, and then feed into the embedding layer.

S2, dividing the data set;

Divide the data set completed in step S1 into training set, verification set and test set, and then perform normalization processing on the features and target execution data on the training set, verification set and test set respectively, wherein the formula of normalization processing is as follows :

S3. Constructing a fully connected neural network model combined with an embedding layer;

The neural network structure includes an input layer, an embedding layer, a connection layer, a hidden layer, and an output layer. The input layer is used to connect input variables, the embedding layer is used to process the classification variables passed from the input layer, and the connection layer is responsible for connecting the variables processed by the embedding layer. features and other features, the hidden layer is composed of multiple neurons for feature extraction and learning, and the last layer of the neural network is the output layer. Among them, the embedding layer, hidden layer and output layer all contain parameters that need to be learned. Label encoding and one-hot encoding as shown in Figure 3 are traditional methods for dealing with categorical variables. In this structure, a general fully connected neural network is combined with an embedding layer to process categorical variables more effectively. If only a general form of fully connected neural network framework is used, this type of structure cannot effectively handle categorical variables and numerical variables in features, and all features will be regarded as numerical variables, which will inevitably affect the model effect. By connecting the label encoding, one-hot encoding embedding layer and the general form of fully connected neural network, it can help the neural network to better learn the information transmitted by the three categorical variables of weather station number, month and season, and also overcome the problem of only The disadvantages of the traditional way of using one-hot encoding or label encoding to deal with categorical variables.

S4. Model hyperparameter setting and model training;

S41. Model hyperparameter setting:

Model hyperparameters are configurations external to the model whose values cannot be estimated from the data. According to specific circumstances, the hyperparameters of the neural network model of the present invention are set as follows: 1) the embedded vector dimension in the embedding layer is 8; 2) the number of hidden layers is 1 layer, and the number of neurons in the hidden layer is 64, and the activation function 3) The number of neurons in the output layer is 1, and the activation function is Linear; the calculation formula of the activation function ReLU is as follows:

where x is a variable.

S42. Model training:

When training the neural network, the loss function is set to mean absolute error (Mean Absolute Error, MAE). In order to prevent the model from overfitting, the regularization method of early stopping is adopted. Set the number of model iterations to 50 times. If the verification error does not decrease but increases for 3 consecutive times during training, stop the training. After the model is trained, then apply the model to the verification set, calculate the model error on the verification set, and verify by comparison Set for hyperparameter selection to obtain the best model, and save the best-performing model on the validation set.

S5. Model evaluation and daily maximum temperature forecast;

To evaluate how good the model is, it is evaluated on the test set. The test set is a sample of data separate from the training and validation sets. The specific steps are: load the saved and trained model, input the characteristic data of the test set in step S2 into the model, and the model can output the predicted value of the daily maximum temperature. By comparing the predicted value output by the model with the actual observed value, various Test method, which can evaluate the forecasting performance of the model independently of the training set.

If the model of the present invention is to be applied to a new sample for daily maximum temperature prediction, the new sample data should be preprocessed according to step S1, and then the processed data can be input into the model to obtain the predicted value of the maximum temperature.

The daily maximum temperature forecast by the present invention from January 1 to December 31, 2020 is compared with the daily maximum temperature forecast of the integrated forecast system IFS of the European Center for Medium-Range Forecasting, as shown in Figure 4 and Figure 5. In the figure, it is realized as a diagonal line, the dotted line is a fitting line, and the shaded part represents the estimated value of the kernel density. It can be seen from the figure that the forecast of the present invention is closer to the observed value (closer to the forecast of the IFS model) than the forecast of the IFS model. diagonal). Compared with the root mean square error (2.746) of the IFS model forecast, the average root mean square error (1.433) of the daily maximum temperature forecast of the present invention in 2020 is reduced by 47.82%, and the absolute average error is reduced by 46.5%.

The above descriptions are only preferred embodiments of the present invention, and it should be understood that the present invention is not limited to the forms disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and Modifications can be made within the scope of the ideas described herein, by virtue of the above teachings or skill or knowledge in the relevant art. However, changes and changes made by those skilled in the art do not depart from the spirit and scope of the present invention, and should all be within the protection scope of the appended claims of the present invention.

Claims (7)

1. A forecasting method based on fully connected neural network daily maximum temperature, characterized in that: said forecasting method comprises: Step 1, preprocessing the observation data and numerical forecast data; Step 2. Divide the preprocessed data into training set, verification set and test set, and perform normalization processing on the features and target execution data on the training set, verification set and test set respectively; Step 3. Construct a fully connected neural network model combined with an embedding layer, and set the hyperparameters of the model and train the model; Step 4. Evaluate the model after training, and predict the maximum temperature through the model. 2. A kind of forecasting method based on fully connected neural network daily maximum air temperature according to claim 1, is characterized in that: described observation data and numerical forecast data are preprocessed and include the following content: S11. Perform data cleaning on the observation data and generate the target file: eliminate the wrong or missing sites, select the maximum temperature observation data of the site with a complete time series, and create the target data set; S12. Read and convert the data format of the historical numerical model: read the GRIB format data of the numerical forecast model, unify the horizontal resolution of the ground layer forecast data and the upper-altitude layer forecast data, and linearly interpolate a small number of missing forecast values in the time dimension The method performs interpolation, and outputs the historical numerical forecast data in nc format; S13. Calculate the upper-altitude combination data of the numerical model: based on the upper-altitude layer data of the numerical model, through the calculation formula

Calculate the eddy advection, through the calculation formula

Calculate the temperature advection, where v represents the wind speed,

is the Laplace operator on the isobaric surface, ζ is the relative vorticity, f is the planetary vorticity, T is the temperature; S14, interpolating the data on the numerical model grid point to the weather station site: adopting the bilinear interpolation method to interpolate the ground data, high-altitude data and high-altitude combined data on the numerical model grid point to the ground weather station site, and interpolating The final numerical model data and target data are output in one file; S15. Build model input features: According to the forecast experience, select the predictors related to the daily maximum temperature forecast on the ground layer and the upper layer, and select the predictors related to the forecast quantity as the model features through the correlation analysis method and the mutual information value method . 3. A kind of forecasting method based on fully connected neural network daily maximum air temperature according to claim 2, is characterized in that: described by correlation analysis method and mutual information value method, selects the predictor relevant with predictor as model feature include: Correlation analysis method: through the correlation coefficient formula

Select a factor with a correlation coefficient greater than 0.3 and reach a significance level of 0.05, where X is the observed value, representing the target, Y is the predicted value, representing the feature, cov(X, Y) is the covariance of X, Y, σ _X is the standard deviation of X, μ _X is the expected E(X) of X; Mutual information value method: By calculating the mutual information value between the target and the feature, the nonlinear correlation between the feature and the target is measured, and the joint distribution of two random variables (X, Y) is set to p(x, y), and the marginal distribution They are p(x), p(y), and the mutual information I(X; Y) is the relative entropy of the joint distribution p(x, y) and the marginal distribution p(x), p(y), and the mutual information value is expressed as

The time lag variable is selected as the maximum temperature observation data with a lag of N days and N+i days, and the auxiliary variable is selected to identify the time information of different stations and samples in the data set to complete the creation of the feature data set. 4. A kind of forecasting method based on fully connected neural network daily maximum air temperature according to claim 3, is characterized in that: described auxiliary variable comprises season, month, meteorological station number, station longitude, station latitude and station altitude; And the station longitude, station latitude, and station altitude in the auxiliary variables are used as numerical variables, and the weather station number, season and month in the auxiliary variables are used as categorical variables. 5. a kind of forecasting method based on fully connected neural network daily maximum temperature according to claim 4, is characterized in that: the fully connected neural network model of the combination embedding layer of construction comprises input layer, embedding layer, connection layer, hidden layer and the output layer; the input layer includes label encoding and one-hot encoding, and the weather station number and month in the classification variables are input into the label encoding, and the seasons in the classification variables are input into the one-hot encoding; the embedding layer uses It is used to process the categorical variables passed by the label encoding in the input layer; the connection layer is used to connect the features processed by the embedding layer, the one-hot encoded features and other features; the hidden layer is composed of multiple neurons for feature extraction And learning; finally output through the output layer. 6. A kind of forecasting method based on fully connected neural network daily maximum temperature according to claim 1, is characterized in that: described hyperparameter setting of model and model training comprise: Set the embedding vector dimension in the embedding layer to 8, the number of hidden layers to 1, the number of neurons in the hidden layer to 64, the activation function to ReLU, the number of neurons in the output layer to 1, and the activation function is Linear; Set the loss function to mean absolute error, use this regularization method to prevent the model from overfitting, set the number of model iterations to 50 times, if the verification error does not decrease but increases for 3 consecutive times during training, stop training, and the model training is complete , then apply the model to the validation set, calculate the model error on the validation set, perform hyperparameter selection by comparing the validation set to obtain the best model, and save the model with the best performance on the validation set. 7. A kind of forecasting method based on fully connected neural network daily maximum temperature according to claim 1, is characterized in that: described model is evaluated to the end of training, and predicting maximum temperature by model comprises: Load the model that has been saved after training, input the characteristic data in the test set in step S2 into the model to output the predicted value of daily maximum temperature, and compare the predicted value output by the model with the actual observed value to make the model independent of the forecast outside the training set performance evaluation; When there is a new data sample to predict the daily maximum temperature, the new data sample is preprocessed according to step S1, and then the processed data is input into the model to obtain the predicted value of the maximum temperature.

Images