Meteorological element sequence prediction-oriented model and method

Abstract

The invention discloses a meteorological element sequence prediction-oriented model and method. The method comprises the following steps: a data input module normalizes input data and performs preprocessing operation; the feature extraction module extracts periodic features and long-term features from the meteorological element data matrix by using two-dimensional convolution; the periodic rule extraction module maps the periodic characteristics to a periodic rule; the long-term trend extraction module maps the long-term features to a long-term trend; the short-term trend extraction module is used for generating meteorological elements to predict the short-term trends of the feedback interval and the prediction interval; the preliminary prediction module adds the periodic rule, the long-term trend and the short-term trend of the meteorological elements to obtain a preliminary prediction result; the noise feedback module takes out prediction of a feedback interval from the preliminary prediction of the meteorological elements, performs subtraction on the prediction and a real feedback sequence to obtain a feedback noise sequence, and models the noise sequence of the prediction interval through the linear module to correct the preliminary prediction; and the final prediction module adds the noise sequence and the preliminary prediction result and performs inverse normalization to obtain a final prediction result. The method has high adaptability.

Description

Technical Field

[0001] The present invention relates to the field of meteorological element prediction, and in particular to a model and method for meteorological element sequence prediction, which is used for sequence prediction of temperature, humidity, air pressure and the like in meteorological elements. Background Art

[0002] Temperature, humidity, and air pressure are important meteorological elements that have a profound impact on society, economy, and environment. Accurate weather forecasts are of great significance in agriculture, energy, water resources management, disaster warning, and other fields. Thanks to the development of the times and technological progress, functional models based on machine learning and deep learning continue to emerge. For example, neural network models can effectively capture the nonlinear relationship between meteorological data, and deep models can process large-scale data and extract hidden features, thereby improving the accuracy and reliability of temperature forecasts.

[0003] However, in the process of time series forecasting and weather forecasting, existing forecasting models have some defects, such as: (1) The internal structure and parameters of deep learning models are uninterpretable, resulting in low credibility and acceptability in practical applications; (2) Machine learning and deep learning models rely on large amounts of data for training, while meteorological data are often inaccurate or incomplete; (3) Excellent models benefit from high model complexity, but are also prone to overfitting and high computational costs; (4) Due to the complexity and uncertainty of the meteorological system, it is difficult for the model to perform well under a variety of different meteorological conditions, and its adaptability needs to be improved.

[0004] Based on the above analysis, it is of great significance to study a model that can effectively balance prediction accuracy and efficiency, model complexity and interpretability. Summary of the invention

[0005] The purpose of the present invention is to provide a model and method for meteorological element sequence prediction in order to address the balance between the accuracy and speed of deep prediction models, the complexity and interpretability of models, and the problem of missing radiation observation data at some observation stations.

[0006] The technical solution adopted by the present invention is:

[0007] A model for meteorological element sequence prediction, including:

[0008] Data input module: obtains target meteorological element data as input data and performs preprocessing operations on the input data to form a meteorological element data matrix;

[0009] Feature extraction module: extract periodic features and long-term features from the meteorological element data matrix using two-dimensional convolution;

[0010] Periodic law extraction module: maps the extracted periodic features to obtain the periodic law of the corresponding meteorological elements;

[0011] Long-term trend extraction module: Map the extracted long-term features to obtain the long-term trend of the corresponding meteorological elements;

[0012] Short-term trend extraction module: generates forecasts of short-term trends of the corresponding meteorological elements for the feedback interval and the prediction interval;

[0013] Preliminary forecast module: sum up the periodic laws, long-term trends and short-term trends of corresponding meteorological elements to obtain preliminary forecast results;

[0014] Noise feedback module: The prediction of the feedback interval is taken from the preliminary prediction of the corresponding meteorological element, and the difference between the actual feedback sequence is obtained to obtain the feedback noise sequence, and then the noise sequence of the prediction interval is modeled through the linear module to correct the preliminary prediction;

[0015] Final prediction module: add up the noise sequence and preliminary prediction results, and then perform inverse normalization to obtain the final prediction result, that is, the final prediction result is composed of four parts: periodic law, long-term trend, short-term trend, and noise.

[0016] Furthermore, the target meteorological element is temperature.

[0017] Furthermore, the feature extraction module is composed of two convolutional layers, two dropout layers, two batch normalization layers, one activation layer and one residual connection. The convolutional layer uses a 3*3 convolution kernel with a step size of 1, and padding is used to keep the input and output dimensions consistent. The residual connection has only one convolutional layer, using a 1*1 convolution kernel with a step size of 1.

[0018] Furthermore, the convolution layer is used to extract features of the input data, and generates a feature map by sliding the convolution kernel on the input data and calculating the dot product with the input data; the discard layer randomly discards (i.e. sets to zero) the output of a part of the neurons in the neural network during the training process to reduce overfitting; the batch normalization layer standardizes the data to have a fixed mean and variance, which helps to accelerate the training process and enhance the generalization ability of the model; the activation layer is used to introduce nonlinearity so that the neural network can learn and represent complex data relationships; the residual connection directly connects the input data to the output of the subsequent layer to form a residual block, which is used to help solve the gradient disappearance and gradient explosion problems in deep neural networks.

[0019] Furthermore, the periodic law extraction module, the long-term trend extraction module and the short-term trend extraction module have the same linear structure, called MLPBlock, which consists of two fully connected layers, two discard layers, one activation layer and one residual connection.

[0020] Furthermore, each input node of the fully connected layer is connected to all output nodes of the layer, and the output is calculated through a weight matrix and a bias term, aiming to perform a linear transformation on the input features.

[0021] A prediction method for a meteorological element sequence comprises the following steps:

[0022] Step 1, data input: obtain the time series data set RX of the target meteorological element, and use the Z-Score normalization method to normalize the data set to obtain the normalized data set and the historical sequence X 1:l =[x1,x2,…,x l ] and fold the historical sequence processing into a matrix Among them, n is the number of samples, d is the number of feature elements; l is the length of the historical sequence; s is the matrix X 1:p The period length of p is the matrix X 1:p The number of cycles;

[0023] Step 2, feature extraction: Get the initial input matrix X 1:p After expanding the channel dimension, we get the expanded input matrix After feature extraction by ConvBlocks, the first output feature map is obtained c is the output channel, and the number of channels increases with the superposition of ConvBlock; the feature map Perform dimensionality conversion (Rearrange) to aggregate the channel dimension c and the cycle number dimension p to obtain the second output feature map matrix cp = c × p;

[0024] Step 3, periodic law extraction: the second output feature map matrix The channel cycle number dimension cp is mapped to one dimension to aggregate different historical cycles and different channels, and the last dimension is compressed (Squeeze) to obtain the periodic law

[0025] Step 4, long-term trend extraction: Two MLPBlocks are used for step-by-step mapping. The first MLPBlock outputs the first feature map The channel dimension c and cycle dimension s are mapped to one dimension to aggregate different historical cycles and channels, and the last dimension is compressed (Squeeze) to obtain the historical long-term trend The second MLPBlock infers the future long-term trend based on the historical long-term trend and obtains the future long-term trend in

[0026] Step 5, short-term trend extraction: construct feedback sequence And x 1:l =[x 1:l-f ,x l-f:l ], f is the length of the feedback sequence, f≤l; the feedback sequence Mapped to the f+h dimension, we get a forecast representing the short-term trend of the feedback interval and the prediction interval.

[0027] Step 6, preliminary forecast: sum up the periodic laws, long-term trends and short-term trends of the corresponding meteorological elements to obtain preliminary forecast results;

[0028] Step 7, Noise Feedback: From Preliminary Prediction Get the prediction for the feedback interval With the real feedback sequence x l-f:l Subtract and get the feedback noise sequence Use MLPBlock to map the noise sequence of the feedback interval to h dimensions and obtain the prediction representing the noise sequence of the prediction interval Preliminary forecasts with revised forecast intervals;

[0029] Step 8, final prediction: add the noise sequence and the preliminary prediction, and then perform inverse normalization to obtain the final prediction result of the model corresponding to the meteorological elements

[0030] Furthermore, in step 1, the target meteorological element is temperature; the length of the historical sequence is l=9×24=216 hours, and the cycle length s of the temperature is 24 hours.

[0031] Furthermore, the prediction length in step 4 is h=7×24=168 hours.

[0032] Furthermore, the modeling process of periodic law is formally expressed as:

[0033] X conv =Rearrange(ConvBlocks(Reshape(x 1:l )))

[0034] x period =Squeeze(MLPBlock c×p→1 (X conv ))

[0036] Furthermore, the modeling process of long-term trends is formally expressed as:

[0037]

[0038] Furthermore, step 6 comprises the following steps:

[0039] Step 6-1, periodic law x period Repeated times, and truncate to obtain the prediction of the feedback interval and prediction interval based on the periodic law

[0040] Step 6-2, long-term trend x longTrend After repeating s times, it is flattened and truncated to obtain the forecast of feedback interval and prediction interval based on long-term trend

[0041] Step 6-3, with the short-term trend x sgortTrend Add them together to get the model's initial prediction

[0042] Furthermore, the process of step 6 is formally expressed as:

[0043] x shortTrend =MLPBlock f→f+h (x l-f:l )

[0044]

[0045] Furthermore, the noise feedback process in step 7 can be formally expressed as:

[0046]

[0047] x Noise =MLPBlock f→h (ε)

[0049] Furthermore, the final prediction process in step 8 can be formally expressed as:

[0050]

[0051] Furthermore, MLPBlock consists of two fully connected layers, two dropout layers, one activation layer and one residual connection.

[0052] Furthermore, each input node of the fully connected layer is connected to all output nodes of the layer, and the output is calculated through the weight matrix and bias term to perform a linear transformation on the input features.

[0053] The present invention adopts the above technical solution, and the beneficial effects are as follows: the balance between prediction accuracy, model complexity and interpretability is achieved through periodic decomposition and long-term and short-term trend modeling. The feedback noise sequence can be effectively used to correct the prediction results of the prediction interval, making up for the problem of missing some observation data. After the model is applied, it is not only applicable to periodic data sets, but also can achieve satisfactory results on non-periodic data sets and data sets with different periodic characteristics of different variables, and has strong adaptability. It performs well on actual meteorological data and shows extremely high interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments;

[0055] Figure 1 A schematic diagram of the principle architecture of a model and method for meteorological element sequence prediction according to the present invention;

[0056] Figure 2 It is a structural schematic diagram of the feature extraction structure ConvBlock of the present invention;

[0057] Figure 3 It is a schematic diagram of the structure of the linear structure MLPBlock of the present invention;

[0058] Figure 4 It is a schematic diagram of the performance results of the present invention on an actual temperature data set. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the invention is applied to the temperature prediction of a certain area below, and the prediction length is set to h=7×24=168, then the review length is l=9×24=216, and the cycle length s of the temperature is 24; the technical solution in the embodiments of the present application is clearly and completely described in conjunction with the drawings in the embodiments of the present application.

[0060] like Figures 1 to 4 As shown in one, the present invention discloses a model for meteorological element sequence prediction, comprising:

[0061] Data input module: obtains target meteorological element data as input data and performs preprocessing operations on the input data to form a meteorological element data matrix;

[0062] Feature extraction module: extract periodic features and long-term features from the meteorological element data matrix using two-dimensional convolution;

[0063] Periodic law extraction module: maps the extracted periodic features to obtain the periodic law of the corresponding meteorological elements;

[0064] Long-term trend extraction module: Map the extracted long-term features to obtain the long-term trend of the corresponding meteorological elements;

[0065] Short-term trend extraction module: generates forecasts of short-term trends of the corresponding meteorological elements for the feedback interval and the prediction interval;

[0066] Preliminary forecast module: sum up the periodic laws, long-term trends and short-term trends of corresponding meteorological elements to obtain preliminary forecast results;

[0067] Noise feedback module: The prediction of the feedback interval is taken from the preliminary prediction of the corresponding meteorological element, and the difference between the actual feedback sequence is obtained to obtain the feedback noise sequence, and then the noise sequence of the prediction interval is modeled through the linear module to correct the preliminary prediction;

[0068] Final prediction module: The final prediction result of the corresponding meteorological element is obtained by adding up the noise sequence of the corresponding meteorological element and the preliminary prediction result, that is, the final prediction result is composed of four parts: periodic law, long-term trend, short-term trend and noise.

[0069] Furthermore, the feature extraction module is composed of two convolutional layers, two dropout layers, two batch normalization layers, one activation layer and one residual connection. The convolutional layer uses a 3*3 convolution kernel with a step size of 1, and padding is used to keep the input and output dimensions consistent. The residual connection has only one convolutional layer, using a 1*1 convolution kernel with a step size of 1.

[0070] Furthermore, the convolution layer is used to extract features of the input data, and generates a feature map by sliding the convolution kernel on the input data and calculating the dot product with the input data; the discard layer randomly discards (i.e. sets to zero) the output of a part of the neurons in the neural network during the training process to reduce overfitting; the batch normalization layer standardizes the data to have a fixed mean and variance, which helps to accelerate the training process and enhance the generalization ability of the model; the activation layer is used to introduce nonlinearity so that the neural network can learn and represent complex data relationships; the residual connection directly connects the input data to the output of the subsequent layer to form a residual block, which is used to help solve the gradient disappearance and gradient explosion problems in deep neural networks.

[0071] Furthermore, the periodic law extraction module, the long-term trend extraction module and the short-term trend extraction module have the same linear structure, called MLPBlock, which consists of two fully connected layers, two discard layers, one activation layer and one residual connection.

[0072] Furthermore, each input node of the fully connected layer is connected to all output nodes of the layer, and the output is calculated through a weight matrix and a bias term, aiming to perform a linear transformation on the input features.

[0073] A method for meteorological element sequence prediction, comprising the following steps:

[0074] Step 1, data input: obtain the temperature series data set RX of the meteorological station in the area to be predicted, and use the Z-Score normalization method to normalize the data set to obtain the normalized data set and the history sequence x 1:1 =[x1,x2,…,x1], and fold the historical sequence processing into a matrix Among them, n is the number of samples, d is the number of feature elements; l is the length of the historical sequence; s is the matrix X 1:p The period length of p is the matrix X 1:p In the temperature series prediction embodiment, usually about one year of historical data is needed, then the data sequence length n is 365*24 with an interval of 1 hour, d represents the number of meteorological stations in a specific area, which is generally greater than 3, and can form a mutual reference of the temperature series to enhance the robustness of the model. l is set to 216, and the cycle of temperature change is 24 hours, so the cycle length s is set to 24, and p is set to 9.

[0075] Step 2, feature extraction: Get the initial input matrix X 1:p After expanding the channel dimension, we get the expanded input matrix After feature extraction by ConvBlocks, the first output feature map is obtained c is the output channel, and the number of channels increases with the superposition of ConvBlock; the feature map Perform dimensionality conversion (Rearrange) to aggregate the channel dimension c and the cycle number dimension p to obtain the second output feature map matrix cp = c × p;

[0076] Step 3, periodic law extraction: the second output feature map matrix The channel cycle number dimension cp is mapped to one dimension to aggregate different historical cycles and different channels, and the last dimension is compressed (Squeeze) to obtain the periodic law

[0077] Step 4, long-term trend extraction: Two MLPBlocks are used for step-by-step mapping. The first MLPBlock outputs the first feature map The channel dimension c and cycle dimension s are mapped to one dimension to aggregate different historical cycles and channels, and the last dimension is compressed (Squeeze) to obtain the historical long-term trend The second MLPBlock infers the future long-term trend based on the historical long-term trend and obtains the future long-term trend in

[0078] Step 5, short-term trend extraction: construct feedback sequence And x 1:l =[x 1:l-f ,x l-f:l ], f is the length of the feedback sequence, f≤l; the feedback sequence Mapped to the f+h dimension, we get a forecast representing the short-term trend of the feedback interval and the prediction interval. When the prediction length is h=168, the feedback sequence length is f=192;

[0079] Step 6, preliminary forecast: sum up the periodic laws, long-term trends and short-term trends of the corresponding meteorological elements to obtain preliminary forecast results;

[0080] Step 7, Noise Feedback: From Preliminary Prediction Get the prediction for the feedback interval With the real feedback sequence x l-f:l Subtract and get the feedback noise sequence Use MLPBlock to map the noise sequence of the feedback interval to h dimensions and obtain the prediction representing the noise sequence of the prediction interval Preliminary forecasts with revised forecast intervals;

[0081] Step 8, final prediction: add the noise sequence and the preliminary prediction, then perform inverse normalization to get the final prediction result of the model

[0082] Furthermore, the modeling process of periodic laws and long-term trends is formally expressed as:

[0083] X conv =Rearrange(ConvBlocks(Reshape(x 1:l )))

[0084] x period =Squeeze(MLPBlock c×p→1 (X conv ))

[0085]

[0086] Furthermore, step 6 comprises the following steps:

[0087] Step 6-1, periodic law x period Repeated times, and truncate to obtain the prediction of the feedback interval and prediction interval based on the periodic law

[0088] Step 6-2, long-term trend x longTrend After repeating s times, it is flattened and truncated to obtain the forecast of feedback interval and prediction interval based on long-term trend

[0089] Step 6-3. Compare them with the short-term trend x shortTrend Add them together to get the model's initial prediction

[0090] Furthermore, the process of step 6 is formally expressed as:

[0091] x shortTrend =MLPBlock f→f+h (x l-f:l )

[0092]

[0093] Furthermore, the noise feedback process in step 7 can be formally expressed as:

[0094]

[0095] x Noise =MLPBlock f→h (ε)

[0097] Furthermore, the final prediction process in step 7 can be formally expressed as:

[0098]

[0100] The specific principle of the present invention is described in detail below:

[0101] Given a time series dataset with n samples and d feature elements Given a history sequence X of length l 1:l =[x1,x2,…,x l ], to predict the future sequence X of length h l+1:l+h =[x l+1 ,x l+2 ,…,x l+h ], so the multivariate prediction task goal is to find a mapping F:X 1:l →X l+1:l+h .

[0102] The data input module modifies the encoder input size according to the number of variables of the selected dataset so that the encoder input size and the number of variables of the dataset are consistent.

[0103] Furthermore, the data input module includes a Reshape operation, which transforms the vector x 1:l =[x1,x2,…,x l ]Folded into a matrix

[0104] The feature extraction module uses a two-dimensional convolution ConvBlock for feature extraction. Let s be the cycle length. The input sequence has a total of p cycles, p = l / s and l%s = 0. Input the matrix X in the feature extraction module 1:p In order to perform convolution calculations, the channel dimension needs to be expanded. After feature extraction by ConvBlocks, the output feature map is obtained c is the output channel, and the number of channels increases with the superposition of ConvBlock. Rearrange, to aggregate the channel dimension c and the cycle number dimension p, to obtain the input matrix cp=c×p

[0105] Furthermore, the feature extraction module is composed of two convolutional layers, two dropout layers, two batch normalization layers, one activation layer and one residual connection. The convolutional layer uses a 3*3 convolution kernel with a step size of 1, and padding is used to keep the input and output dimensions consistent. The residual connection has only one convolutional layer, using a 1*1 convolution kernel with a step size of 1.

[0106] Furthermore, the convolution layer is used to extract features of the input data. It generates a feature map by sliding the convolution kernel over the input data and calculating the dot product between it and the input data. The dropout layer randomly drops (i.e. sets to zero) the output of a portion of neurons in the neural network during training to reduce overfitting. The batch normalization layer normalizes the data to have a fixed mean and variance, which helps to speed up the training process and enhance the generalization ability of the model. The activation layer is used to introduce nonlinearity so that the neural network can learn and represent complex data relationships.

[0107] Furthermore, the residual connection directly connects the input data to the output of the subsequent layer to form a residual block, which is used to help solve the gradient disappearance and gradient explosion problems in deep neural networks.

[0108] The periodic law extraction module, the long-term trend extraction module and the short-term trend extraction module have the same linear structure, called MLPBlock, which consists of two fully connected layers, two discard layers, one activation layer and one residual connection.

[0109] Furthermore, each input node of the fully connected layer is connected to all output nodes of the layer, and the output is calculated through a weight matrix and a bias term, aiming to perform a linear transformation on the input features.

[0110] In the periodic law extraction module, the channel cycle number dimension cp is mapped to one dimension to aggregate different historical cycles and different channels, and the last one dimension is squeezed to obtain the periodic law.

[0111] The long-term trend extraction module uses two MLPBlocks for step-by-step mapping. The first MLPBlock maps the channel dimension c and the period dimension s to one dimension to aggregate different historical periods and channels, and squeezes the last dimension to obtain the historical long-term trend. The second MLPBlock infers the future long-term trend based on the historical long-term trend and obtains the future long-term trend in

[0112] Furthermore, let the length of the feedback sequence be f, then the feedback sequence And x 1:l =[x 1:l-f ,x l-f:l ], f≤l. The modeling process of the above periodic law and long-term trend can be formally expressed as:

[0113] X conv =Rearrange(ConvBlocks(Reshape(x 1:l )))

[0114] x period =Squeeze(MLPBlock c×p→1 (X conv ))

[0115]

[0116] The short-term trend extraction module will feedback the sequence Mapped to the f+h dimension, representing the prediction of short-term trends in the feedback interval and prediction interval

[0117] Next, we sum up the above cyclical patterns, long-term trends, and short-term trends to get a preliminary forecast, which includes the following steps:

[0118] S1, periodic law x period Repeated times, and truncate to obtain the prediction of the feedback interval and prediction interval based on the periodic law

[0119] S2, long-term trend x longTrend After repeating s times, it is flattened and truncated to obtain the forecast of feedback interval and prediction interval based on long-term trend

[0120] S3. Compare them with the short-term trend x shortTrend Add them together to get the model's initial prediction

[0121] Furthermore, the above process can be formally expressed as:

[0122] x shortTrend =MLPBlock f→f+j (x l-f:l )

[0123]

[0124] In the noise prediction module, from the preliminary prediction Get the prediction for the feedback interval With the real feedback sequence x l-f:l Subtract and get the feedback noise sequence Then, MLPBlock is used to map the noise sequence of the feedback interval to the h dimension, which represents the prediction of the noise sequence of the prediction interval. To correct the prediction interval. Add the noise sequence and the preliminary prediction to get the final prediction result of the model.

[0125] Furthermore, this process can be formally expressed as:

[0126]

[0127] x Noise =MLPBlock f→h (ε)

[0128]

[0129] The present invention adopts the above technical solution, and the beneficial effects are as follows: the balance between prediction accuracy, model complexity and interpretability is achieved through periodic decomposition and long-term and short-term trend modeling. The feedback noise sequence can be effectively used to correct the prediction results of the prediction interval, making up for the problem of missing some observation data. After the model is applied, it is not only applicable to periodic data sets, but also can achieve satisfactory results on non-periodic data sets and data sets with different periodic characteristics of different variables, and has strong adaptability. It performs well on actual meteorological data and shows extremely high interpretability.

[0130] Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians of the art without making creative work are within the scope of protection of the present application.

Claims

1. A model for forecasting meteorological element sequences, characterized by: It includes: Data input module: obtains target meteorological element data as input data and performs preprocessing operations on the input data to form a meteorological element data matrix; Feature extraction module: extract periodic features and long-term features from the meteorological element data matrix using two-dimensional convolution; Periodic law extraction module: maps the extracted periodic features to obtain the periodic law of the corresponding meteorological elements; Long-term trend extraction module: Map the extracted long-term features to obtain the long-term trend of the corresponding meteorological elements; Short-term trend extraction module: generates forecasts of short-term trends of the corresponding meteorological elements for the feedback interval and the prediction interval; Preliminary prediction module: sum up the periodic laws, long-term trends and short-term trends of the corresponding meteorological elements to obtain the preliminary prediction results; Noise feedback module: take the prediction of the feedback interval from the preliminary prediction of the corresponding meteorological elements and make a difference with the real feedback sequence to obtain the feedback noise sequence, and then use the linear module to model the noise sequence of the prediction interval to correct the preliminary prediction; Final prediction module: The final prediction result of the corresponding meteorological element is obtained by adding up the noise sequence of the corresponding meteorological element and the preliminary prediction result, that is, the final prediction result is composed of four parts: periodic law, long-term trend, short-term trend and noise.

2. A model for meteorological element sequence prediction according to claim 1, characterized in that: The feature extraction module consists of two convolutional layers, two dropout layers, two batch normalization layers, one activation layer and one residual connection. The convolutional layer uses a 3*3 convolution kernel with a step size of 1, and padding is used to keep the input and output dimensions consistent. The residual connection has only one convolutional layer, using a 1*1 convolution kernel with a step size of 1.

3. A model for meteorological element sequence prediction according to claim 2, characterized in that: The convolution layer is used to extract the features of the input data. It generates a feature map by sliding the convolution kernel on the input data and calculating the dot product between the convolution kernel and the input data. The dropout layer randomly discards the output of some neurons in the neural network during training to reduce overfitting. The batch normalization layer standardizes the data to have a fixed mean and variance. The activation layer is used to introduce nonlinearity so that the neural network can learn and represent complex data relationships. The residual connection directly connects the input data to the output of the subsequent layer to form a residual block.

4. The model for meteorological element sequence prediction according to claim 1, characterized in that: The periodic law extraction module, the long-term trend extraction module and the short-term trend extraction module have the same linear structure called MLPBlock, which consists of two fully connected layers, two discard layers, one activation layer and one residual connection.

5. A model for meteorological element sequence prediction according to claim 4, characterized in that: Each input node of the fully connected layer is connected to all output nodes of the layer, and the output is calculated through the weight matrix and bias term to perform a linear transformation on the input features.

6. A method for meteorological element sequence prediction, using a meteorological element sequence prediction model according to any one of claims 1 to 5, characterized in that: The method comprises the following steps: Step 1, data input: obtain the time series dataset of the target meteorological elements and the historical sequence X 1:l =[x1,x2,…,x l ] and fold the historical sequence processing into a matrix Among them, n is the number of samples, d is the number of characteristic elements; l is the length of the historical sequence; s is the matrix X 1:p The period length of p is the matrix X 1:p The number of cycles; Step 2, feature extraction: Get the initial input matrix X of meteorological elements 1:p After expanding the channel dimension, we get the expanded input matrix After feature extraction by ConvBlocks, the first output feature map is obtained c is the output channel, and the number of channels increases with the superposition of ConvBlock; the feature map Perform dimensionality conversion (Rearrange) to aggregate the channel dimension c and the cycle number dimension p to obtain the second output feature map matrix cp = c × p; Step 3, periodic law extraction: The second output feature map matrix of meteorological elements The channel cycle number dimension cp is mapped to one dimension to aggregate different historical cycles and different channels, and the last dimension is compressed to obtain the periodic law Step 4, long-term trend extraction: Two MLPBlocks are used for step-by-step mapping. The first MLPBlock converts the first output feature map of the meteorological element The channel dimension c and cycle dimension s are mapped to one dimension to aggregate different historical cycles and channels, and the last dimension is compressed to obtain the historical long-term trend. The second MLPBlock infers the future long-term trend based on the historical long-term trend and obtains the future long-term trend in h is the prediction length; Step 5, short-term trend extraction: construct feedback sequence And x 1:l =[x 1:l-f ,x l-f:l ], f is the length of the feedback sequence, f≤l; the feedback sequence Mapped to the f+h dimension, we get a forecast representing the short-term trend of the feedback interval and the prediction interval. Step 6, preliminary forecast: sum up the periodic laws, long-term trends and short-term trends of the corresponding meteorological elements to obtain the preliminary forecast results. Step 7, Noise Feedback: From Preliminary Prediction Get the prediction for the feedback interval With the real feedback sequence x l-f:l Subtract and get the feedback noise sequence Use MLPBlock to map the noise sequence of the feedback interval to h dimensions and obtain the prediction representing the noise sequence of the prediction interval Preliminary forecasts with revised forecast intervals; Step 8, final prediction: add up the noise sequence and the preliminary prediction, perform inverse normalization to obtain the final prediction result of the model corresponding to the meteorological elements 7. A method for meteorological element sequence prediction according to claim 6, characterized in that: In step 1, the target meteorological element is temperature; the cycle length s of temperature is 24 hours, and the length of the historical sequence is l = 9 × 24 = 216 hours.

8. The method for meteorological element sequence prediction according to claim 6, characterized in that: The prediction length in step 4 is h = 7 × 24 = 168 hours.

9. The method for meteorological element sequence prediction according to claim 6, characterized in that: Step 6 contains the following steps: Step 6-1, periodic law x period Repeated times, and truncate to obtain the prediction of the feedback interval and prediction interval based on the periodic law Step 6-2, long-term trend x longTrend After repeating s times, it is flattened and truncated to obtain the forecast of feedback interval and prediction interval based on long-term trend Step 6-3. Compare them with the short-term trend x shortTrend Add them together to get the model's initial prediction 10. The method for meteorological element sequence prediction according to claim 6, characterized in that: MLPBlock consists of two fully connected layers, two discard layers, one activation layer and one residual connection; each input node of the fully connected layer is connected to all the output nodes of the layer, and the output is calculated through the weight matrix and bias term to perform a linear transformation on the input features.

Images