A neural network-based regional meteorological model error control method

Abstract

The application discloses a regional meteorological model error control method based on a neural network, and particularly relates to the field of regional meteorological model error control, and comprises the following steps: obtaining reanalysis data of a first data source and station observation data of a second data source, and forming an input matrix containing spatial coordinates, a time index and meteorological element values; performing a preprocessing operation on the input matrix, wherein the preprocessing comprises: interpolating and correcting missing values, normalizing data of different dimensions, and performing sequence alignment based on a unified time index to form a first input matrix and a second input matrix with consistent structures. By constructing a double residual error channel and performing residual reciprocal correction in the training process, dynamic balance fusion of the reanalysis data and the station observation data is realized in the time and space scales, so that the cross-source data fusion imbalance problem in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of regional meteorological model error control technology, and more specifically, to a method for regional meteorological model error control based on neural networks. Background Technology

[0002] In meteorological forecasting for wind and solar power generation, a common practice is to train neural networks using both reanalysis data and field observation data. Reanalysis data has the advantage of broad coverage, providing a large-scale overall climate background, but its information is relatively smooth and it is difficult to reflect sudden local changes. In contrast, field observation data can capture more subtle fluctuations in wind speed and irradiance, but due to limitations in spatial coverage and equipment stability, its performance is often highly localized and incomplete. Current technologies typically process these two types of data by simply normalizing or directly concatenating them into the model, lacking a targeted approach to addressing the differences between them. As a result, during training, neural networks tend to gradually favor the type of data that is easier to fit, either over-relying on the overall trend of the reanalysis data or over-relying on the high-frequency details of the field data. Consequently, the important information from the other type of data is weakened during prediction, and the model exhibits a significant bias. This bias accumulates over long-term operation, eventually reducing the model's ability to respond to key meteorological changes, especially when wind speed changes suddenly or irradiance drops sharply, the prediction results may show significant deviations. This leads to the core problem: existing neural networks lack an effective mechanism to balance the differences between the two types of data when fusing reanalysis data and field observation data, causing bias in the model during training and prediction, thus resulting in imbalance in cross-source data fusion. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a regional meteorological model error control method based on neural networks. By constructing dual residual channels and performing residual reciprocal correction during training, the reanalysis data and station observation data are dynamically balanced and fused at the spatiotemporal scale, thereby solving the cross-source data fusion imbalance problem mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a regional meteorological model error control method based on neural networks, comprising: S1. Obtain the reanalysis data from the first data source and the station observation data from the second data source, and form an input matrix containing spatial coordinates, time index and meteorological element values; S2. Perform preprocessing operations on the input matrix. The preprocessing includes: interpolating and correcting missing values, normalizing data of different dimensions, and performing sequence alignment based on a unified time index to form a first input matrix and a second input matrix with consistent structure. S3. Construct a neural network structure containing two residual channels, input the first input matrix into the first residual channel, input the second input matrix into the second residual channel, and solve the first residual vector and the second residual vector in each channel respectively; S4. During the training process, a residual reciprocal correction operation is performed. The residual reciprocal correction operation includes: generating a first correction term based on the first residual vector and applying it to the second residual vector; simultaneously generating a second correction term based on the second residual vector and applying it to the first residual vector; thereby forming a first residual and a second residual after reciprocal correction in the bidirectional correction loop. S5. Input the first residual and the second residual after mutual correction to the prediction layer. Generate the prediction sequence of regional meteorological elements in the prediction layer. Based on the balance state of the residual in the bidirectional correction cycle, output the final prediction result after error control.

[0005] In a preferred embodiment, S1 includes: S1-1. Extract reanalysis data from the first data source and solve for the corresponding longitude coordinates, latitude coordinates, time index, and element values ​​of wind speed, temperature, humidity, and irradiance in the reanalysis data to construct the first dataset. S1-2. Extract station observation data from the second data source, and extract the station geographic coordinates, observation time index, and element values ​​of wind speed, temperature, humidity, and irradiance to construct the second dataset; S1-3. Perform synchronization operation on the first dataset and the second dataset based on the time index, and match the corresponding meteorological element values ​​under the same or adjacent spatial coordinates to generate a unified data alignment result. S1-4. Arrange the spatial coordinates, time indices, and meteorological element values ​​obtained from the data alignment results according to the row and column index rules to form an input matrix.

[0006] In a preferred embodiment, S2 includes: S2-1. Identify the location of missing values ​​in the input matrix and perform linear interpolation calculation based on the meteorological element values ​​corresponding to adjacent time indices. When the missing interval is greater than the preset threshold, use cubic spline interpolation to solve for the meteorological element values ​​of the missing points. S2-2. After interpolation correction, the meteorological element values ​​are statistically analyzed according to the element category, and interval scaling is performed based on the maximum and minimum values ​​to generate a normalized meteorological element value sequence. S2-3. Align the normalized input matrix according to a unified time index, and perform corresponding arrangement of meteorological element values ​​of reanalysis data and station observation data under the same time index. When a certain time index is missing an observation value, fill it with the interpolation result of the previous step. S2-4. The sequence-aligned data is reorganized according to the dimensional order of spatial coordinates, time index, and meteorological element values ​​to form a first input matrix and a second input matrix with consistent structure, which are then used as inputs for subsequent residual calculations.

[0007] In a preferred embodiment, in S3, the process of constructing a neural network structure containing dual residual channels includes: S3-1. Establish a first residual channel and a second residual channel in the neural network, allocate the first input matrix formed by the reanalysis data to the first residual channel, and allocate the second input matrix formed by the station observation data to the second residual channel. S3-2. In the first residual channel, a convolution operation is performed on the first input matrix. Local features are extracted in the spatial and temporal dimensions by sliding convolution kernels. Then, a nonlinear activation operation is performed to maintain the nonlinear relationship between features. At the output end, the convolution output is added to the original input matrix by residual superposition operation. Then, the difference calculation is performed on the addition result and the first input matrix to solve for the first residual vector. S3-3. Perform convolution operation on the second input matrix in the second residual channel. Extract local features in the spatial and temporal dimensions by sliding convolution kernel. Then perform nonlinear activation operation to maintain the nonlinear relationship between features. At the output end, add the convolution output to the original input matrix by residual superposition operation. Then perform difference calculation on the addition result and the second input matrix to solve for the second residual vector.

[0008] In a preferred embodiment, in S3, the process of constructing the neural network structure containing dual residual channels further includes: S3-4. Perform dimension preservation operations on the solved first residual vector and second residual vector respectively: align the two in spatial coordinate dimension, time index dimension and meteorological element dimension, and perform normalization operation to make the first residual vector and the second residual vector consistent in numerical scale, forming residual output.

[0009] In a preferred embodiment, in S4, the process of performing residual reciprocal correction during training includes: S4-1. Perform scale alignment operation on the first residual vector and the second residual vector, count the maximum and minimum values ​​of each vector and perform interval scaling to make the two vectors consistent in numerical range and dimensional structure, forming an aligned residual pair. S4-2. Generate the first corrected weight sequence based on the first residual vector: Perform a difference operation on the first residual vector at the time index to obtain the rate of change sequence, and generate component weights according to the product of the residual magnitude and the rate of change. S4-3. Generate the second correction weight sequence based on the second residual vector: Perform a difference operation on the second residual vector at the time index to obtain the rate of change sequence, and generate component weights by multiplying the residual magnitude and the rate of change. S4-4. Generate the first correction term and the second correction term and perform bidirectional operation: Scale the second residual vector component by component using the first correction weight sequence and perform weighted difference operation to obtain the first correction term. At the same time, scale the first residual vector component by component using the second correction weight sequence and perform weighted difference operation to obtain the second correction term. Apply the first correction term to the second residual vector and apply the second correction term to the first residual vector to form an updated residual vector pair. S4-5. Execute a bidirectional correction loop, repeating S4-2 to S4-4, and calculate the residual consistency index after each loop. The residual consistency index is the average of the absolute differences between the first residual vector and the second residual vector on their corresponding components. When the residual consistency index is less than a preset lower limit or the number of loops reaches a set upper limit, terminate the loop and output the first residual and the second residual after mutual correction.

[0010] In a preferred embodiment, in S5, the process of inputting the first residual and the second residual after reciprocal correction together into the prediction layer and outputting the final prediction result includes: The first residual vector and the second residual vector after mutual inversion correction are concatenated point by point in the time index dimension and combined in parallel in the spatial coordinate dimension to form the residual input tensor. The residual input tensor is input to the prediction layer, where multi-head attention is first performed to identify the dependency between different spatial coordinates and time indices and to generate attention-weighted residual feature sequences. The residual feature sequence is input into a fully connected neural network layer. In the fully connected neural network layer, a weighted summation operation and a nonlinear activation operation are performed at each time index position to generate a preliminary prediction vector sequence of meteorological elements. The preliminary prediction vector sequence is input into the recursive calculation unit, where state updates and output solutions are performed sequentially based on time indexes to form a prediction sequence of regional meteorological elements. After the prediction sequence is generated, the prediction sequence is combined with the residual balance state in the bidirectional correction loop, the residual weight distribution is calculated, and the prediction sequence is corrected with the residual weight distribution to output the final prediction result after error control.

[0011] The technical effects and advantages of this invention are as follows: In this scheme, the reanalysis data and the station observation data are balanced by using dual residual channels and reciprocal correction mechanism, which avoids the model’s over-reliance on a single data source during training and fundamentally eliminates the prediction bias problem caused by the imbalance of cross-source data fusion. By interpolating and correcting missing values, performing interval scaling on multidimensional data, and aligning sequences based on a unified time index during the preprocessing stage, this scheme effectively ensures the structural consistency of data from different sources, providing a reliable data foundation for subsequent residual calculation and correction. By combining convolution extraction, nonlinear activation and residual superposition, this scheme forms a closed loop of multidimensional feature extraction and differential calculation in the dual residual channels, enabling the residual vector to accurately represent the differences and local perturbation features of different data sources. Through a bidirectional residual reciprocal correction loop, this scheme utilizes the weight terms generated by difference to achieve dynamic balance correction of the residuals, gradually reducing the deviation between the two types of residual vectors during training, thereby improving the stability and reliability of the model under multi-source data fusion. By introducing a multi-head attention mechanism, a combination of fully connected computation and recursive computation units in the prediction layer, this scheme fully captures the spatiotemporal dependencies of meteorological elements in the prediction stage, and combines the residual balance state to weighted correct the prediction sequence, thereby achieving high-precision error control of regional meteorological elements. Attached Figure Description

[0012] Figure 1 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Refer to the instruction manual appendix Figure 1 An embodiment of the present invention provides a method for error control of a regional meteorological model based on a neural network, comprising: S1. Obtain the reanalysis data from the first data source and the station observation data from the second data source, and form an input matrix containing spatial coordinates, time index and meteorological element values; S2. Perform preprocessing operations on the input matrix. The preprocessing includes: interpolating and correcting missing values, normalizing data of different dimensions, and performing sequence alignment based on a unified time index to form a first input matrix and a second input matrix with consistent structure. S3. Construct a neural network structure containing two residual channels, input the first input matrix into the first residual channel, input the second input matrix into the second residual channel, and solve the first residual vector and the second residual vector in each channel respectively; S4. During the training process, a residual reciprocal correction operation is performed. The residual reciprocal correction operation includes: generating a first correction term based on the first residual vector and applying it to the second residual vector; simultaneously generating a second correction term based on the second residual vector and applying it to the first residual vector; thereby forming a first residual and a second residual after reciprocal correction in the bidirectional correction loop. S5. Input the first residual and the second residual after mutual correction to the prediction layer. Generate the prediction sequence of regional meteorological elements in the prediction layer. Based on the balance state of the residual in the bidirectional correction cycle, output the final prediction result after error control.

[0015] S1 includes: S1-1. Extract reanalysis data from the first data source and solve for the corresponding longitude coordinates, latitude coordinates, time index, and element values ​​of wind speed, temperature, humidity, and irradiance in the reanalysis data to construct the first dataset. S1-2. Extract station observation data from the second data source, and extract the station geographic coordinates, observation time index, and element values ​​of wind speed, temperature, humidity, and irradiance to construct the second dataset; S1-3. Perform synchronization operation on the first dataset and the second dataset based on the time index, and match the corresponding meteorological element values ​​under the same or adjacent spatial coordinates to generate a unified data alignment result. S1-4. Arrange the spatial coordinates, time indices, and meteorological element values ​​obtained from the data alignment results according to the row and column index rules to form an input matrix; It should be noted that spatial coordinates refer to the longitude and latitude coordinates obtained from the reanalysis data, as well as the geographical coordinates of the station obtained from the station observation data. The time index refers to the time index obtained from the reanalysis data and the observation time index obtained from the station observation data; Meteorological element values ​​refer to the element values ​​of wind speed, temperature, humidity, and irradiance extracted from reanalysis data and station observation data; In addition, in this scheme, the reanalysis data refers to the regional meteorological dataset calculated by numerical weather models in combination with global multi-source observation data. This regional meteorological dataset includes longitude coordinates, latitude coordinates, time index, and values ​​of wind speed, temperature, humidity, and irradiance. The station observation data refers to the local meteorological data collected in real time by the corresponding sensor equipment in the wind and solar stations. This dataset includes the station's geographical coordinates, observation time index, and values ​​of wind speed, temperature, humidity, and irradiance. Secondly, although the two types of data share the same meteorological indicators, the reanalysis data is used to provide continuous background information over a regional scale, while the station observation data is used to provide the actual measurement results within a local area. Since the reanalysis data has smoothing errors within a local area, and the station observation data has problems with insufficient coverage and missing measurements, this scheme uses both types of data simultaneously: the reanalysis data is used to ensure the continuity of the large-scale climate background, and the station observation data is used to supplement and correct local disturbances. By introducing these two types of data in parallel into the input matrix, the accumulation of bias caused by a single data source can be avoided, and complementary constraints can be formed in the subsequent residual correction process.

[0016] S2 includes: S2-1. Identify the location of missing values ​​in the input matrix and perform linear interpolation calculation based on the meteorological element values ​​corresponding to adjacent time indices. When the missing interval is greater than the preset threshold, use cubic spline interpolation to solve for the meteorological element values ​​of the missing points. S2-2. After interpolation correction, the meteorological element values ​​are statistically analyzed according to the element category, and interval scaling is performed based on the maximum and minimum values ​​to generate a normalized meteorological element value sequence. S2-3. Align the normalized input matrix according to a unified time index, and perform corresponding arrangement of meteorological element values ​​of reanalysis data and station observation data under the same time index. When a certain time index is missing an observation value, fill it with the interpolation result of the previous step. S2-4. The sequence-aligned data is reorganized according to the dimensional order of spatial coordinates, time index, and meteorological element values ​​to form a first input matrix and a second input matrix with consistent structure, which are then used as inputs for subsequent residual calculations. It should be noted that cubic spline interpolation involves constructing a set of cubic polynomials between known data points, ensuring that each polynomial is continuous within adjacent intervals and that the function values, first derivative, and second derivative are continuous at the endpoints of the intervals, thus forming a smooth curve. The calculation method includes: establishing a piecewise cubic polynomial equation system and solving the coefficient matrix in combination with boundary conditions to obtain the interpolation function in all intervals, thereby calculating the corresponding value at any missing point. In S2-2, interval scaling refers to linearly transforming a set of original data according to its maximum and minimum values, so that its numerical range is remapped to a specified interval, including 0 to 1 or −1 to 1. Its significance is to eliminate the numerical imbalance between different meteorological elements caused by different units and value ranges, so that the neural network can process the input on the same scale during training, and improve the comparability and stability of each element in the calculation. In addition, the functional forms of linear transformation in interval scaling include: linear mapping functions with fixed slope and intercept, which directly map any input value to the target interval by subtracting the minimum value from the original value and then scaling it up or down proportionally by the difference between the maximum and minimum values.

[0017] In S3, the process of constructing a neural network structure containing dual residual channels includes: S3-1. Establish a first residual channel and a second residual channel in the neural network, allocate the first input matrix formed by the reanalysis data to the first residual channel, and allocate the second input matrix formed by the station observation data to the second residual channel. S3-2. In the first residual channel, a convolution operation is performed on the first input matrix. Local features are extracted in the spatial and temporal dimensions by sliding convolution kernels. Then, a nonlinear activation operation is performed to maintain the nonlinear relationship between features. At the output end, the convolution output is added to the original input matrix by residual superposition operation. Then, the difference calculation is performed on the addition result and the first input matrix to solve for the first residual vector. S3-3. Perform convolution operation on the second input matrix in the second residual channel. Extract local features in the spatial and temporal dimensions by sliding convolution kernel. Then perform nonlinear activation operation to maintain the nonlinear relationship between features. At the output end, add the convolution output to the original input matrix by residual superposition operation. Then perform difference calculation on the addition result and the second input matrix to solve for the second residual vector.

[0018] In S3, the process of constructing a neural network structure with dual residual channels also includes: S3-4. Perform dimension preservation operations on the solved first residual vector and second residual vector respectively: align the two in spatial coordinate dimension, time index dimension and meteorological element dimension, and perform normalization operation to make the first residual vector and the second residual vector consistent in numerical scale, forming residual output that can be used in subsequent correction steps.

[0019] In S4, the process of performing residual reciprocal correction during training includes: S4-1. Perform scale alignment operation on the first residual vector and the second residual vector, count the maximum and minimum values ​​of each vector and perform interval scaling to make the two vectors consistent in numerical range and dimensional structure, forming an aligned residual pair. S4-2. Generate the first corrected weight sequence based on the first residual vector: Perform a difference operation on the first residual vector at the time index to obtain the rate of change sequence, and generate component weights according to the product of the residual magnitude and the rate of change. S4-3. Generate the second correction weight sequence based on the second residual vector: Perform a difference operation on the second residual vector at the time index to obtain the rate of change sequence, and generate component weights by multiplying the residual magnitude and the rate of change. S4-4. Generate the first correction term and the second correction term and perform bidirectional operation: Scale the second residual vector component by component using the first correction weight sequence and perform weighted difference operation to obtain the first correction term. At the same time, scale the first residual vector component by component using the second correction weight sequence and perform weighted difference operation to obtain the second correction term. Apply the first correction term to the second residual vector and apply the second correction term to the first residual vector to form an updated residual vector pair. S4-5. Execute a bidirectional correction loop, repeating S4-2 to S4-4, and calculate the residual consistency index after each loop. The residual consistency index is the average of the absolute differences between the first residual vector and the second residual vector on their corresponding components. When the residual consistency index is less than a preset lower limit or the number of loops reaches a set upper limit, terminate the loop and output the first residual and the second residual after mutual correction.

[0020] In S5, the process of inputting the first and second residuals, after mutual inversion correction, into the prediction layer and outputting the final prediction result includes: The first residual vector and the second residual vector after mutual inversion correction are concatenated point by point in the time index dimension and combined in parallel in the spatial coordinate dimension to form the residual input tensor. The residual input tensor is input into the prediction layer. In the prediction layer, multi-head attention operation is first performed to identify the dependency relationship between different spatial coordinates and temporal indices and generate attention-weighted residual feature sequences. The multi-head attention operation refers to mapping the input sequence into query vectors, key vectors, and value vectors. The attention weight is obtained by calculating the similarity between the query vector and the key vector. Then, the value vector is weighted and summed. Different heads perform the operation in parallel in different subspaces. Finally, the results of each head are concatenated to generate the weighted feature sequence. The residual feature sequence is input into a fully connected neural network layer. In this layer, a weighted summation and nonlinear activation operation are performed at each time index position to generate a preliminary prediction vector sequence of meteorological elements. The fully connected neural network layer means that each neuron in this layer is connected to all feature nodes input from the previous layer, with each connection corresponding to a weight value. Each neuron also contains a bias term. During computation, the input features are multiplied and accumulated one by one with their corresponding weights, then the bias term is added, and finally, the output value of the neuron is generated through a nonlinear activation function. Nonlinear activation operation refers to introducing a function with nonlinear mapping characteristics after linear transformations such as fully connected layers or convolutions, converting the linear combination result into a nonlinear output, thereby enhancing the neural network's ability to fit complex relationships. Common nonlinear activation functions include: the Rectified Linear Unit (ReLU), whose output is the upper limit of the input value and zero; the Sigmoid function, whose output smoothly varies between zero and one; and the Hyperbolic Tangent (Tanh), whose output varies between negative one and one. The preliminary prediction vector sequence is input into the recursive calculation unit, where state updates and output solutions are performed sequentially based on time indexes to form a prediction sequence of regional meteorological elements. After the prediction sequence is generated, the prediction sequence is combined with the residual balance state in the bidirectional correction loop, the residual weight distribution is calculated, and the prediction sequence is corrected with the residual weight distribution to output the final prediction result after error control. It should be noted that in the recursive calculation unit, the residual input tensor after reciprocal correction is first input sequentially according to the time index. The input vector corresponding to each time index and the hidden state passed down from the previous time index are used together in the calculation. Next, the input vector and the hidden state are multiplied by the weight matrix and added. After adding the bias term, a new hidden state is generated through a nonlinear activation function. This hidden state numerically carries the feature information of the historical time and the current time. Then, the new hidden state is multiplied by the output weight and the output bias is added. After passing through the activation function, the meteorological element output vector corresponding to the time index is obtained. Finally, the output vectors obtained from all time index positions are concatenated in chronological order to form a regional meteorological element prediction sequence containing spatial coordinate dimension, time index dimension, and meteorological element dimension. Furthermore, the residual balance state refers to the degree to which the first and second residual vectors tend to be close in terms of time indices and spatial coordinates after the bidirectional correction cycle is completed. It includes two parts: first, the absolute difference between the two vectors at each component, reflecting the degree of residual deviation of the corresponding meteorological element; second, the average of this absolute difference across all components, reflecting the overall consistency level. In calculation, the absolute difference between the first and second residual vectors at the same index position is first calculated component-by-component to obtain the difference vector. Then, the average of this difference vector is calculated to obtain the consistency index. Subsequently, based on the difference vector in the residual balance state, the components of the prediction sequence are weighted according to their magnitude: positions with smaller differences are assigned higher weights, indicating strong consistency between the two types of data; positions with larger differences are assigned lower weights, indicating uncertainty. The residual weight distribution generated in this way can perform weighted correction of the prediction sequence in space and time based on residual consistency, thereby obtaining the prediction result after error control.

[0021] It should be noted that this scheme first extracts spatial coordinates, time indices, and meteorological element values ​​such as wind speed, temperature, humidity, and irradiance from reanalysis data and station observation data, respectively, and forms a unified input matrix through synchronous matching and matrix processing; then, preprocessing operations are performed on the input matrix to sequentially complete the interpolation correction of missing values, interval scaling and normalization of data with different dimensions, and sequence alignment based on the unified time index, thereby generating a first input matrix and a second input matrix with consistent structure; During the modeling phase, the two types of input matrices are fed into the dual residual channels of the neural network. Spatiotemporal features are extracted through convolution operations, nonlinear activation preserves the relationship between elements, and residual superposition and differential decomposition are performed to obtain the first residual vector and the second residual vector, respectively. During training, residual reciprocal correction is performed on the two residual vectors. Specifically, a weight sequence is generated by difference and a correction term is constructed to form a bidirectional cyclic correction until the residual consistency index converges, thereby obtaining the residual output after reciprocal correction. In the prediction phase, the corrected residual inputs are concatenated and fed into the prediction layer. First, a multi-head attention mechanism identifies the spatiotemporal dependencies. Then, the prediction sequence of regional meteorological elements is gradually calculated through a fully connected layer and a recursive computation unit. Finally, the prediction sequence is weighted and corrected by combining the weight distribution generated by the residual balance state, and the final prediction result with error control is output.

[0022] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for error control of a regional meteorological model based on a neural network, characterized in that, include: S1. Obtain the reanalysis data from the first data source and the station observation data from the second data source, and form an input matrix containing spatial coordinates, time index and meteorological element values; S2. Perform preprocessing operations on the input matrix. The preprocessing includes: interpolating and correcting missing values, normalizing data of different dimensions, and performing sequence alignment based on a unified time index to form a first input matrix and a second input matrix with consistent structure. S3. Construct a neural network structure containing two residual channels, input the first input matrix into the first residual channel, input the second input matrix into the second residual channel, and solve the first residual vector and the second residual vector in each channel respectively; S4. During the training process, a residual reciprocal correction operation is performed. The residual reciprocal correction operation includes: generating a first correction term based on the first residual vector and applying it to the second residual vector; simultaneously generating a second correction term based on the second residual vector and applying it to the first residual vector; thereby forming a first residual and a second residual after reciprocal correction in the bidirectional correction loop. In S4, the process of performing residual reciprocal correction during training includes: S4-1. Perform scale alignment operation on the first residual vector and the second residual vector, count the maximum and minimum values ​​of each vector and perform interval scaling to make the two vectors consistent in numerical range and dimensional structure, forming an aligned residual pair. S4-2. Generate the first corrected weight sequence based on the first residual vector: Perform a difference operation on the first residual vector at the time index to obtain the rate of change sequence, and generate component weights according to the product of the residual magnitude and the rate of change. S4-3. Generate the second correction weight sequence based on the second residual vector: Perform a difference operation on the second residual vector at the time index to obtain the rate of change sequence, and generate component weights by multiplying the residual magnitude and the rate of change. S4-4. Generate the first correction term and the second correction term and perform bidirectional operation: Scale the second residual vector component by component using the first correction weight sequence and perform weighted difference operation to obtain the first correction term. At the same time, scale the first residual vector component by component using the second correction weight sequence and perform weighted difference operation to obtain the second correction term. Apply the first correction term to the second residual vector and apply the second correction term to the first residual vector to form an updated residual vector pair. S4-5. Execute a bidirectional correction loop, repeating S4-2 to S4-4, and calculate the residual consistency index after each loop. The residual consistency index is the average of the absolute differences between the first residual vector and the second residual vector on their corresponding components. When the residual consistency index is less than a preset lower limit or the number of loops reaches a set upper limit, terminate the loop and output the first residual and the second residual after mutual correction. S5. Input the first residual and the second residual after mutual correction to the prediction layer. Generate the prediction sequence of regional meteorological elements in the prediction layer. Based on the balance state of the first residual and the second residual in the bidirectional correction cycle, output the final prediction result after error control. In S5, the process of inputting the first and second residuals, after mutual inversion correction, into the prediction layer and outputting the final prediction result includes: The first residual vector and the second residual vector after mutual inversion correction are concatenated point by point in the time index dimension and combined in parallel in the spatial coordinate dimension to form the residual input tensor. The residual input tensor is input to the prediction layer, where multi-head attention is first performed to identify the dependency between different spatial coordinates and time indices and to generate attention-weighted residual feature sequences. The residual feature sequence is input into a fully connected neural network layer. In the fully connected neural network layer, a weighted summation operation and a nonlinear activation operation are performed at each time index position to generate a preliminary prediction vector sequence of meteorological elements. The preliminary prediction vector sequence is input into the recursive calculation unit, where state updates and output solutions are performed sequentially based on time indexes to form a prediction sequence of regional meteorological elements. After the prediction sequence is generated, the prediction sequence is combined with the residual balance state in the bidirectional correction loop, the residual weight distribution is calculated, and the prediction sequence is corrected with the residual weight distribution to output the final prediction result after error control.

2. The method for error control of a regional meteorological model based on a neural network according to claim 1, characterized in that: S1 includes: S1-1. Extract reanalysis data from the first data source and solve for the corresponding longitude coordinates, latitude coordinates, time index, and element values ​​of wind speed, temperature, humidity, and irradiance in the reanalysis data to construct the first dataset. S1-2. Extract station observation data from the second data source, and extract the station geographic coordinates, observation time index, and element values ​​of wind speed, temperature, humidity, and irradiance to construct the second dataset; S1-3. Perform synchronization operation on the first dataset and the second dataset based on the time index, and match the corresponding meteorological element values ​​under the same or adjacent spatial coordinates to generate a unified data alignment result. S1-4. Arrange the spatial coordinates, time indices, and meteorological element values ​​obtained from the data alignment results according to the row and column index rules to form an input matrix.

3. The method for error control of a regional meteorological model based on a neural network according to claim 2, characterized in that: S2 includes: S2-1. Identify the location of missing values ​​in the input matrix and perform linear interpolation calculation based on the meteorological element values ​​corresponding to adjacent time indices. When the missing interval is greater than the preset threshold, use cubic spline interpolation to solve for the meteorological element values ​​of the missing points. S2-2. After interpolation correction, the meteorological element values ​​are statistically analyzed according to the element category, and interval scaling is performed based on the maximum and minimum values ​​to generate a normalized meteorological element value sequence. S2-3. Align the normalized input matrix according to a unified time index, and perform corresponding arrangement of meteorological element values ​​of reanalysis data and station observation data under the same time index. When a certain time index is missing an observation value, fill it with the interpolation result of the previous step. S2-4. The sequence-aligned data is reorganized according to the dimensional order of spatial coordinates, time index, and meteorological element values ​​to form a first input matrix and a second input matrix with consistent structure, which are then used as inputs for subsequent residual calculations.

4. The method for error control of a regional meteorological model based on a neural network according to claim 3, characterized in that: In S3, the process of constructing a neural network structure containing dual residual channels includes: S3-1. Establish a first residual channel and a second residual channel in the neural network, allocate the first input matrix formed by the reanalysis data to the first residual channel, and allocate the second input matrix formed by the station observation data to the second residual channel. S3-2. In the first residual channel, a convolution operation is performed on the first input matrix. Local features are extracted in the spatial and temporal dimensions by sliding convolution kernels. Then, a nonlinear activation operation is performed to maintain the nonlinear relationship between features. At the output end, the convolution output is added to the original input matrix by residual superposition operation. Then, the difference calculation is performed on the addition result and the first input matrix to solve for the first residual vector. S3-3. Perform convolution operation on the second input matrix in the second residual channel. Extract local features in the spatial and temporal dimensions by sliding convolution kernel. Then perform nonlinear activation operation to maintain the nonlinear relationship between features. At the output end, add the convolution output to the original input matrix by residual superposition operation. Then perform difference calculation on the addition result and the second input matrix to solve for the second residual vector.

5. The method for error control of a regional meteorological model based on a neural network according to claim 4, characterized in that: In S3, the process of constructing a neural network structure with dual residual channels also includes: S3-4. Perform dimension preservation operations on the solved first residual vector and second residual vector respectively: align the two in spatial coordinate dimension, time index dimension and meteorological element dimension, and perform normalization operation to make the first residual vector and the second residual vector consistent in numerical scale, forming residual output.

Images