A single-station hourly air temperature prediction method based on data fusion and hybrid convolution

Abstract

The present application relates to a kind of single-station hourly temperature prediction methods based on data fusion and mixed convolution, belong to meteorological prediction field.The present application selects important single-station meteorological observation element, obtains single-station historical observation sequence;Extract multiple-element prediction data, obtain numerical prediction space-time sequence;Data normalization;A three-dimensional convolution module is constructed, input after numerical prediction space-time sequence of normalization, finally produce time series containing spatial features;Build interactive learning model ICM, to single-station historical observation sequence and the time series extracted by three-dimensional convolution processing, generate the time characteristic information of two kinds of data;Build MCNN network model, fuse the two branch time characteristic information obtained in S3, after again interactive learning and fully connected, form prediction result;Temperature prediction is carried out using the model.The time series prediction model built by the present application improves stability and prediction accuracy, and the method is suitable for single-station 72-120 hours of hourly temperature prediction.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting, specifically involving a method for single-station hourly temperature forecasting based on data fusion and hybrid convolution. Background Technology

[0002] Meteorological station data represents historical or current changes in meteorological elements and can be used to predict atmospheric conditions at that station. However, it is difficult to obtain the characteristics of nonlinear atmospheric variation curves. Numerical weather prediction aims to simulate the atmospheric prediction process and provide future weather conditions. However, numerical prediction is based on discrete numerical calculations, and the results are often grid data with errors, making it difficult to obtain accurate local weather forecasts. In addition, the complex and variable local weather conditions in mountainous areas pose challenges to forecasting. Therefore, it is possible to combine numerical prediction data with historical data, giving full play to their respective advantages, to conduct atmospheric prediction at specific stations. Summary of the Invention

[0003] (a) Technical problems to be solved

[0004] The technical problem this invention aims to solve is how to provide a single-station hourly temperature forecasting method based on data fusion and hybrid convolution, in order to address the challenges posed by numerical forecasting, which is based on discrete numerical calculations and often results in grid data with errors, making it difficult to obtain accurate local weather forecasts. Furthermore, the complex and changeable local weather conditions in mountainous areas pose a challenge to forecasting.

[0005] (II) Technical Solution

[0006] To address the aforementioned technical problems, this invention proposes a single-station hourly temperature forecasting method based on data fusion and hybrid convolution, which includes the following steps:

[0007] S1. Data Preprocessing:

[0008] S11. Select important meteorological observation elements for individual stations and obtain historical observation sequences for those stations;

[0009] S12. Extract multi-element forecast data to obtain the numerical forecast spatiotemporal sequence;

[0010] S13, Data Normalization;

[0011] S2, 3D Convolutional Feature Extraction:

[0012] Construct a 3D convolutional module, input the normalized numerical forecast spatiotemporal sequence, and finally produce a time series containing spatial features;

[0013] S3, Interactive Learning Model (ICM):

[0014] An interactive learning model ICM is constructed. For single-station historical observation sequences and time series extracted by three-dimensional convolution, a one-dimensional convolutional extraction network CNN1D is used to extract the local correlation of the time series. A hierarchical structure is established to extract sequences at different time scales, generating time feature information of the two types of data.

[0015] S4, Hybrid Convolutional Network Model MCNN:

[0016] An MCNN network model is constructed, which integrates the temporal feature information of the two branches obtained in S3, and then forms a prediction result through interactive learning and full connection; the MCNN network model is trained.

[0017] S5. Result Prediction:

[0018] Based on the trained MCNN network model, the latest observation data from the site and numerical model forecast data are used. After preprocessing, the data are imported into the model to output hourly temperature forecast data for 72-120 hours.

[0019] (III) Beneficial Effects

[0020] This invention proposes a single-station hourly temperature forecasting method based on data fusion and hybrid convolution. Specifically, it presents a time-series prediction model using MCNN (Mixed Convolution Neural Network), which effectively utilizes historical single-station data and numerical forecast data for hourly temperature prediction, improving stability and forecast accuracy. This method is applicable to hourly temperature forecasts for single stations ranging from 72 to 120 hours. Attached Figure Description

[0021] Figure 1 The structure of the three-dimensional convolutional model (CNN) of this invention;

[0022] Figure 2 This is the structure of the interactive convolutional module (ICM) of the present invention;

[0023] Figure 3 This invention relates to the one-dimensional convolution extraction network cnn1d.

[0024] Figure 4 This is the overall structure of the spatiotemporal convolutional network of the present invention;

[0025] Figure 5 Example of prediction results for the test set;

[0026] Figure 6 For comparison of prediction results;

[0027] Figure 7 For comparison of prediction results, see section two;

[0028] Figure 8 The prediction results are compared with those of the third comparison. Detailed Implementation

[0029] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0030] This invention relates to the field of hourly temperature forecasting technology in meteorology, and is a method for single-station medium-term hourly temperature forecasting based on data fusion and hybrid convolution.

[0031] The technical problem to be solved by this invention is to design a model architecture based on deep neural networks, integrate the information features of station data and forecast data, and realize hourly temperature prediction for a single station for 72-120 hours.

[0032] Meteorological station data represents historical or current changes in meteorological elements and can be used to predict atmospheric conditions at that station. However, it is difficult to obtain the characteristics of nonlinear atmospheric variation curves. Numerical weather prediction aims to simulate the atmospheric prediction process and provide future weather conditions. However, numerical prediction is based on discrete numerical calculations, and its results are often grid data with errors, making it difficult to obtain accurate local weather predictions. In addition, the complex and variable local weather conditions in mountainous areas pose challenges to forecasting. Therefore, it is possible to fuse numerical prediction data and historical data to give full play to their respective advantages for station atmospheric prediction. This invention proposes a time series prediction model based on data fusion and mixed convolution neural network (MCNN), which can effectively use single-station historical data and numerical prediction data to predict hourly temperature, improving stability and prediction accuracy. This method is applicable to hourly temperature prediction of 72-120 hours for a single station.

[0033] This invention designs a deep neural network MCNN based on hybrid convolution, builds a spatial convolution module and a time series module, and achieves hourly temperature prediction for a single station within 120 hours through feature fusion, providing the prediction result in one go.

[0034] The technical solution adopted in this invention is as follows: a method for hourly temperature forecasting of a single station in mountainous areas based on data fusion and hybrid convolution, comprising the following steps:

[0035] S1. Data Preprocessing:

[0036] S11. Select important meteorological observation elements for individual stations and obtain historical observation sequences for those stations;

[0037] S12. Extract multi-element forecast data to obtain the numerical forecast spatiotemporal sequence;

[0038] S13, Data Normalization.

[0039] S2, 3D Convolutional Feature Extraction (CNN):

[0040] Construct a 3D convolution module containing a series of spatial convolutions, downsampling, activation functions, fully connected layers, etc., input the normalized numerical forecast spatiotemporal sequence, and finally produce a time series containing spatial features.

[0041] S3, Interactive Learning (ICM):

[0042] For single-station historical observation sequences and time series extracted by three-dimensional convolution, a one-dimensional convolution filter is used to extract the local correlation of the time series, and a hierarchical structure is established to extract sequences at different time scales, generating time feature information of the two types of data.

[0043] S4, Hybrid Convolutional Network (MCNN):

[0044] The two branch time feature information obtained from S3 are fused together, and a prediction result is formed through another interactive learning and fully connected processing.

[0045] S5. Result Prediction:

[0046] Based on the trained model, the latest station data and numerical model forecast data are used, preprocessed, imported into the model, and output hourly temperature forecast data for 72-120 hours.

[0047] Example 1:

[0048] The method for hourly temperature forecasting of single-station mountain areas based on data fusion and hybrid convolution mainly includes: data preprocessing, 3D convolutional feature extraction, interactive learning, spatiotemporal convolutional network, and prediction result output.

[0049] S1. Data Preprocessing:

[0050] S11. Select important single-station meteorological observation elements

[0051] The selected meteorological observation elements include: temperature, humidity, air pressure, wind speed, wind direction, and total cloud cover, to obtain the single-station historical observation sequence S:

[0052]

[0053] Where L is the time series length and N is the number of features. Observational data from the station over a past period (e.g., L = 240 hours) are used as input. Before this, to reduce model complexity, importance screening is required. A random forest method is used for the temperature feature; noise is added to each of the other features to determine the importance of that feature. The greater the loss, the more critical the feature to which noise is added. The results are shown in Table 1, using air temperature as the target.

[0054] Table 1. Importance scores of variables

[0055]

[0056] S12. Extracting multi-element forecast data

[0057] Numerical model forecast data was extracted from European fine-grid data centered on this station to obtain spatial data in time series representation. The data spatial range is 5°, the accuracy is 0.125°, and the time interval is 3 or 6 hours. Forecast data features are diverse; the MCNN model of this invention uses 2-meter air temperature, surface air pressure, surface temperature, total cloud cover, and topographic data as selected features to form a forecast dataset, which is recorded as a numerical forecast spatiotemporal sequence P:

[0058]

[0059] Where L' represents the length of the time series, N' represents the number of meteorological elements, and W×H represents the size of the spatial grid.

[0060] S13, Data Normalization

[0061] The zero-mean normalization method is used to normalize the two types of data in S11 and S12.

[0062] S'=(SS m ) / S s (3)

[0063] P'=(PP m ) / P s (4)

[0064] Where S' and P' are the normalized values, and S and P are the values ​​before normalization, S m Let S be the mean of S. s Let S be the standard deviation, and P be the standard deviation of S. m Let P be the mean of P. s Let P be the standard deviation.

[0065] S2, 3D Convolutional Feature Extraction:

[0066] Convolutional neural networks have played a crucial role in computer vision. A convolutional kernel has a local receptive field, which can capture local features and perform spatial downsampling to continuously extract more detailed features. For forecast grid data, convolutional operations can be performed to extract meteorological element information from the forecast data. This invention utilizes the local receptive field and spatial downsampling of the convolutional kernel to construct a module to extract time-series feature information from forecast data near a station and match it with historical data.

[0067] Figure 1The structure of the three-dimensional convolutional model is shown. A three-dimensional convolutional module (CNN) is constructed, which includes a series of spatial convolutions, downsampling, activation functions, fully connected components, etc., and finally forms a time series containing spatial features. The input of the convolutional module is the normalized numerical forecast spatiotemporal sequence P', and the output is a time series P1 containing spatial feature information.

[0068] For this site, this paper selects 81×81 grid data of fine European grids with a time frame of 0-120 hours as the spatiotemporal information of the vicinity of the site, i.e., P in step S12 after normalization of P'. Multi-layer convolutional operations are used to extract features from the grid data in the time series, and after downsampling, LeakyReLU activation is used. Finally, the output is expanded into one-dimensional information through a fully connected layer.

[0069] The 3D convolutional CNN network used in this invention includes: 1. A 3D spatial convolutional layer Conv3d(5,32,3), increasing the number of channels to 32; 2. A 3D spatial convolutional layer Conv3d(32,64,3), increasing the number of channels to 64; 3. A first spatial downsampling layer MaxPool3d(3), reducing the grid width and height by a factor of 3; 4. A 3D spatial convolutional layer Conv3d(64,32,3), reducing the number of channels to 32; 5. A 3D spatial convolutional layer Conv3d(32,16,3), reducing the number of channels to 64. The number of channels is reduced to 16; 6. The second spatial downsampling layer MaxPool3d(3) reduces the grid width and height by a factor of 3 again; 7. The three-dimensional spatial convolutional layer Conv3d(16,4,3) reduces the number of channels to 4; 8. The three-dimensional spatial convolutional layer Conv3d(4,1,3) reduces the number of channels to 1; 9. The first activation function layer LeakyRelu(); 10. The first fully connected layer Linear(); 11. The second activation function layer LeakyRelu(); 12. The second fully connected layer Linear().

[0070] S3, Interactive Learning Model (ICM):

[0071] An interactive learning model ICM is constructed. For single-station historical observation sequences and time series extracted by three-dimensional convolution, a one-dimensional convolutional extraction network cnn1d is used to extract the local correlation of the time series, and a hierarchical structure is established to extract sequences at different time scales.

[0072] The detailed structure of the ICM model is as follows: Figure 2The input time series is layer 1. The input time series is first split into sequence pairs based on parity, forming layer 2. Then, the sequences in layer 2 are further split into sequence pairs based on parity, and so on. The processing of each small sequence pair is denoted as an interactive convolutional block (ICB), performing interactive learning. In the final layer, the outputs of all ICBs in this layer are concatenated and then passed through a fully connected layer to obtain the output of the ICM model.

[0073] Each small sequence is split and downsampled to generate one subsequence pair. The process is as follows:

[0074] X odd =X 2i-1 ,i=1,2,…,L / 2 (5)

[0075] X even =X 2i ,i=1,2,…,L / 2 (6)

[0076] Where X represents the initial sequence as a single-station historical observation sequence S' or a layer-by-layer split sequence (Split) of the time series P1 output by numerical weather prediction in step S2. odd X even This represents the downsampled subsequence pairs, namely the even-numbered sequence and the odd-numbered sequence. Downsampling the original sequence along the time dimension allows the module to study dynamic information at different time resolutions.

[0077] The interactive convolutional block (ICB) processing steps for each subsequence pair are as follows: 1) Process the subsequence X... odd and X even One-dimensional convolutional extraction networks (CNN1D) are used respectively, and then converted into the form of exp. Each is then multiplied by the atomic sequence of the other, as shown in formula (7), to obtain and 2) Then, through repeated one-dimensional convolutional extraction networks, the... and Projecting to the hidden state, add respectively and As shown in formula (8), the output X' of the ICB module is obtained. odd and X' even .

[0078]

[0079]

[0080] Where σ, ρ and τ are both one-dimensional convolutional extraction networks (CNN1D), such as... Figure 3As shown, the one-dimensional convolutional extraction network includes: one-dimensional padding ReplicationPad1d, one-dimensional convolution Conv1d, LeakyReLU activation, dropout layer, one-dimensional convolution Conv1d, and Tanh activation. Here, T represents the sequence length, C represents the number of channels, and K represents the convolution kernel.

[0081] Finally, after repeated downsampling, the time series X' at different resolutions are updated using a one-dimensional convolutional extraction network cnn1d and interactive convolutional blocks ICB. odd and X' even Then, the subsequences are rearranged through a reverse parity splitting operation and concatenated into a new sequence, which is then added to the original sequence via residual connections. Furthermore, the final output is generated through a connection layer. The specific operations are as follows:

[0082] X' = ​​Revese(X' odd ,X' even (9)

[0083] X out =Linear(X+X') (10)

[0084] Where X' represents the new sequence representation, X out This indicates the final output; Linear indicates a fully connected layer.

[0085] 4. Hybrid Convolutional Network Model (MCNN):

[0086] MCNN is the overall deep neural network of this invention, such as Figure 4 As shown, the specific steps are as follows: 1) Use random forest to normalize the meteorological element S with high contribution, then use S' as input one, and use the normalized P' of the European fine-grid numerical forecast plus topographic data sequence P as input two; 2) Use a 3D convolutional network (CNN) to extract spatial sequence features P1 from input two P'; 3) Use an interactive learning model (ICM) to interactively learn input S' and feature P1 respectively, obtaining Y1 and Y2 respectively, at which point intermediate loss functions L1 and L2 can also be obtained; 4) Project the time series Y1 and Y2 (one-dimensional convolution Conv1d) to obtain two equal-length sequences, then concatenate (cat) to obtain Y3, i.e., (FusionBlock); 5) Pass the output Y3 from the previous step through an interactive convolutional network (ICM) again to obtain Y4; 6) Pass Y4 through a fully connected (Linear) network to form the prediction result Y, at which point the loss function L3 is obtained. The main operations are as follows:

[0087] Y1=ICM(S') (11)

[0088] Y2=ICM(CNN(P')) (12)

[0089] Y3=CAT(α(Y1),β(Y2)) (13)

[0090] Y = Linear(ICM(Y3)) (14)

[0091] Where S' represents input one, P' represents input two, CNN represents a three-dimensional convolutional module, ICM represents an interactive learning model, α and β represent one-dimensional convolution (Conv1d), which is used to adjust the sequence length, CAT represents concatenation and merging, which is used to fuse the features of two branches, and Linear represents fully connected.

[0092] Furthermore, an intermediate-supervision-based loss function is used throughout the training of the MCNN network model to supervise two intermediate generated results: single-station historical data and numerical forecast data. The overall loss function includes the loss function before the fusion of the two branch information, as specifically expressed below:

[0093] L=aL1+bL2+cL3 (15)

[0094] Where L1, L2, and L3 represent the Loss1 loss function, and a, b, and c represent the weight parameters of the three loss functions, respectively. The sum of the weight parameters is 1, and can be set to 0.5, 0.1, and 0.4.

[0095] 5. Result Prediction:

[0096] Based on the trained model, new probe data and numerical forecast data are used, preprocessed, and imported into the model to predict temperature sequences.

[0097] Effect verification

[0098] Figure 5 The example provided is a test set for verification. The blue line represents the measured temperature at a single station, and the red line represents the predicted temperature. The horizontal axis represents the prediction time, and the vertical axis represents the temperature value. Through statistical verification of 145 records over a 120-hour period, the model's average absolute error in predicting hourly temperatures is 1.2°C, and its accuracy is 93% when the error is less than 3°C.

[0099] Figure 6-8 The latest results are used to predict hourly temperatures. The blue line represents the measured temperature at a single station, and the red line represents the predicted temperature. The horizontal axis represents the prediction time, and the vertical axis represents the temperature value. The bar chart represents the prediction difference; red indicates an absolute difference greater than 3. MES is the mean absolute error, and ACC is the accuracy rate. The overall results show that this invention has achieved excellent results.

[0100] This invention presents an hourly temperature forecasting method based on data fusion and hybrid convolution, MCNN (Mixed Convolution Neural Network), which can effectively solve the problem of low forecast accuracy when using only station data or model numerical forecast products, and improve the accuracy of hourly temperature forecasts. Under the condition of an error of 3°, the forecast accuracy can reach 92%.

[0101] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for single-station hourly temperature forecasting based on data fusion and hybrid convolution, characterized in that, The method includes the following steps: S1. Data Preprocessing: S11. Select important meteorological observation elements for individual stations and obtain historical observation sequences for those stations; S12. Extract multi-element forecast data to obtain the numerical forecast spatiotemporal sequence; S13, Data Normalization; S2, 3D Convolutional Feature Extraction: Construct a 3D convolutional module, input the normalized numerical forecast spatiotemporal sequence, and finally produce a time series containing spatial features; S3, Interactive Learning Model (ICM): An interactive learning model ICM is constructed. For single-station historical observation sequences and time series extracted by three-dimensional convolution, a one-dimensional convolutional extraction network CNN1D is used to extract the local correlation of the time series. A hierarchical structure is established to extract sequences at different time scales, generating time feature information of the two types of data. S4, Hybrid Convolutional Network Model MCNN: An MCNN network model is constructed, which integrates the temporal feature information of the two branches obtained in S3, and then forms a prediction result through interactive learning and full connection; the MCNN network model is trained. S5. Result Prediction: Based on the trained MCNN network model, the latest observation data from the site and numerical model forecast data are used. After preprocessing, the data are imported into the model and output hourly temperature forecast data for 72-120 hours. in, In step S4, the MCNN network model includes: 1) using random forest to select meteorological elements with high contribution. After normalization As input one, the European fine-grid numerical weather prediction plus topographic data sequence After normalization As input two; 2) Input two Spatial sequence features are extracted using a 3D convolutional CNN module. 3) Use the interactive learning model ICM to process the input. and spatial sequence features Interactive learning was conducted to obtain... and At this point, the intermediate loss function can also be obtained. and 4) Time series and Perform one-dimensional convolution and Conv1d projection to obtain two equal-length sequences, then concatenate them to obtain CAT. 5) Output from the previous step After another interactive learning model ICM, the result is obtained ;6) The prediction result is generated after the fully connected layer. At this point, the loss function is obtained. Specifically: (11) (12) (13) (14) wherein, represents input one, represents input two, CNN represents a three-dimensional convolution module, ICM represents an interactive learning model, and represents a one-dimensional convolution Conv1d, which adjusts the sequence length, represents concatenation, which fuses the features of two branches, represents a full connection.

2. The single-station hourly temperature forecasting method based on data fusion and hybrid convolution as described in claim 1, characterized in that, Step S11 specifically includes: selecting meteorological observation elements and obtaining single-station historical observation sequences. : （1） in It is the length of the time series. The number of elements is used. For the temperature element, the random forest method is adopted. The importance of the element is determined by adding noise to each of the other elements. The greater the loss, the more critical the element with added noise.

3. The single-station hourly temperature forecasting method based on data fusion and hybrid convolution as described in claim 2, characterized in that, Step S12 specifically includes: extracting numerical model forecast data from the European fine-grid data centered on this station to obtain spatial data in time series representation. The data spatial range is 5°, the accuracy is 0.125°, and the time interval is 3 or 6 hours. Meteorological elements are selected to form a forecast dataset, which is recorded as a numerical forecast spatiotemporal sequence. : (2) in Indicates the length of the time series. Indicates the number of meteorological elements. Indicates the size of the spatial grid.

4. The single-station hourly temperature forecasting method based on data fusion and hybrid convolution as described in claim 3, characterized in that, Step S13 specifically includes: normalizing the two types of data in S11 and S12 using the zero-mean normalization method. (3) (4) in and The value after normalization. and This is the value before normalization. for The mean, for standard deviation for The mean, for The standard deviation.

5. The single-station hourly temperature forecasting method based on data fusion and hybrid convolution as described in claim 4, characterized in that, Step S2 specifically includes: constructing a three-dimensional convolutional module (CNN), comprising: spatial convolution, downsampling, activation function, and fully connected layer. The input to this convolutional module is the normalized numerical prediction spatiotemporal sequence. Multi-layer convolutional operations are used to extract features from gridded time-series data, and then Leaky is applied after downsampling. Relu The activation process is followed by a fully connected layer to expand the output into one-dimensional information, resulting in a time series output containing spatial feature information. .

6. The single-station hourly temperature forecasting method based on data fusion and hybrid convolution as described in claim 5, characterized in that, The 3D convolutional module CNN network selects an 81×81 grid, including: a 3D spatial convolutional layer Conv3d(5,32,3), increasing the number of channels to 32; a 3D spatial convolutional layer Conv3d(32,64,3), increasing the number of channels to 64; a first spatial downsampling layer MaxPool3d(3), reducing the grid width and height by a factor of 3; a 3D spatial convolutional layer Conv3d(64,32,3), reducing the number of channels to 32; and a 3D spatial convolutional layer Conv3d(32,16,32,3), reducing the number of channels to 32. 3) Reduce the number of channels to 16; the second spatial downsampling layer MaxPool3d(3) reduces the grid width and height by a factor of 3 again; the three-dimensional spatial convolutional layer Conv3d(16,4,3) reduces the number of channels to 4; the three-dimensional spatial convolutional layer Conv3d(4,1,3) reduces the number of channels to 1; the first activation function layer LeakyRelu(); the first fully connected layer Linear(); the second activation function layer LeakyRelu(); and the second fully connected layer Linear().

7. The single-station hourly temperature forecasting method based on data fusion and hybrid convolution as described in claim 5 or 6, characterized in that, In step S3, the ICM model includes: the input time series is the first layer; the input time series is first split into a sequence pair according to parity as the second layer; then each sequence in the second layer is divided into a sequence pair according to parity, and so on; the processing of each small sequence pair is denoted as an interactive convolutional block (ICB), interactive learning is performed, and the outputs of all ICBs in the last layer are concatenated and then passed through a fully connected layer to obtain the output of the ICM model.

8. The single-station hourly temperature forecasting method based on data fusion and hybrid convolution as described in claim 7, characterized in that, Each small sequence is split and downsampled to generate one subsequence pair. The process is as follows: (5) (6) Where X represents the initial sequence as a single-station historical observation sequence. Or the layer-by-layer split sequence of the time series P1 output by the numerical weather prediction in step S2; , This represents the downsampled subsequence pairs, namely the even-numbered sequence and the odd-numbered sequence; The ICB processing steps for each subsequence pair are as follows: 1) Process the subsequences... and One-dimensional convolution is used to extract the CNN1D network, and then converted to... The forms are then multiplied by the dot product of the atomic sequences of the other party, as shown in formula (7), to obtain... and ;2) Then, extract the data through the repeated one-dimensional convolutional network CNN1D. and Projecting to the hidden state, add respectively and See formula (8) to obtain the output of the ICB module. and ; (7) (8) in , , , All of these are one-dimensional convolutional extraction networks, cnn1d. The one-dimensional convolutional extraction network cnn1d includes: one-dimensional padding ReplicationPad1d, one-dimensional convolution Conv1d, LeakyReLU activation, dropout layer, one-dimensional convolution Conv1d, and Tanh activation; where T represents the sequence length. Finally, after repeated downsampling, the time series at different resolutions are updated using a one-dimensional convolutional extraction network CNN1D and interactive convolutional blocks (ICB). and Then, the subsequences are rearranged through a reverse parity splitting operation and concatenated into a new sequence, which is then added to the original sequence via residual connections. Furthermore, the final output is generated through a connection layer. The specific operations are as follows: (9) (10) in This represents a new sequence representation. This indicates the final output. This indicates a fully connected layer.

9. The single-station hourly temperature forecasting method based on data fusion and hybrid convolution as described in claim 8, characterized in that, The entire MCNN network model training uses a loss function constructed based on intermediate supervision. The supervision is provided by two intermediate generated results: historical data from a single station and numerical forecast data. The overall loss function includes the loss function before the fusion of information from the two branches, as shown below: (15) in , and express Loss1 loss function , and These represent the weight parameters of the three loss functions, and the sum of the weight parameters is 1.

Images