Overhead transmission line ampacity prediction method based on numerical weather prediction model correction and transfer learning

Abstract

A method for predicting the current carrying capacity (FHC) of overhead transmission lines based on numerical weather prediction model correction and transfer learning includes: acquiring NWP data and historical meteorological data of the overhead transmission lines, and dividing the data into source domain and target domain data; subtracting the source domain NWP data from the historical meteorological data of the overhead transmission lines to obtain an initial tower micro-meteorological data error sequence; using a source domain tower micro-meteorological data error model to correct the NWP data, freezing the model parameters after training, and saving the model; unfreezing the fully connected layer of the source domain tower micro-meteorological data error model, retraining the fully connected layer with target domain error data, fine-tuning the parameters of the source domain tower micro-meteorological data error model, and finally obtaining the predicted values ​​of each meteorological factor in the target domain. The predicted FHC of the overhead transmission lines is solved based on the heat balance equation. This method improves the accuracy of FHC prediction for target overhead transmission lines under the constraint of small sample data.

Description

Technical Field

[0001] This invention relates to the field of overhead transmission line current carrying capacity prediction technology, specifically to a method for predicting overhead transmission line current carrying capacity based on numerical weather prediction model correction and transfer learning. Background Technology

[0002] With the surge in social electricity demand and the large-scale grid connection of renewable energy, increasing the power transmission capacity of overhead transmission lines is urgently needed. The current maximum current (current carrying capacity) of existing overhead transmission lines is calculated based on static current carrying capacity under harsh weather conditions, which is overly conservative, leading to insufficient utilization of the transmission capacity. Dynamic Line Rating (DLR) technology can predict the current carrying capacity of overhead transmission lines based on micro-meteorological conditions, thereby improving their current-carrying capacity. Therefore, to fully tap the power transmission potential of existing overhead transmission lines, it is imperative to develop advanced methods for predicting the current carrying capacity of overhead transmission lines.

[0003] Currently, research methods for predicting the current carrying capacity of overhead transmission lines include:

[0004] (1): Chinese Invention Patent "A Dynamic Current Carrying Capacity Prediction Method for Overhead Transmission Lines Based on RNN Model" (Application No.: 202311319013.0): This patent uses historical data of overhead transmission lines to filter meteorological observation data with a correlation higher than a set threshold with the current carrying capacity data; constructs a sample dataset; and obtains a current carrying capacity prediction model for overhead transmission lines based on the sample dataset and the RNN model. This invention patent uses regional historical meteorological data to predict the current carrying capacity of the line, but the actual meteorological environment on the towers differs from the regional meteorological data. Obviously, errors cannot be avoided when using regional historical meteorological data for prediction.

[0005] (2): Reference [1]: J. Zhan, CY Chung and E. Demeter, Time Series Modeling for Dynamic Thermal Rating of Overhead Lines[J]. IEEE Transactions on Power Systems, vol. 32, no. 3, pp. 2172-2182, May 2017. To reduce DTR fluctuations, a multivariate regression model containing four micro-meteorological factors was first established, then the seasonal and diurnal characteristics were preserved through hourly standardization, and finally the standardized sequence was modeled using an autoregressive integral moving average model; Reference [2]: Jiang JA, Liang YT, Chen CP, et al. On Dispatching Line Ampacities of Power Grids Using Weather-BasedConductor Temperature Forecasts[J]. IEEE Transactions on Smart Grid, vol. 9, no. 1, pp. 406-415, Jan. 2018. Based on principal component regression models, weather data from meteorological stations was used to predict the line's current carrying capacity and determine potential thermal bottlenecks and current carrying capacity margins at different locations along the line. However, the aforementioned traditional statistical prediction models have significant shortcomings in characterizing complex nonlinear relationships and have limited adaptability to rapid changes in meteorological conditions and extreme operating conditions.

[0006] (3): Reference [4]: ​​Wang Kongsen, Sheng Gehao, Sun Xuri, et al. Online prediction of dynamic capacity of transmission line based on radial basis neural network [J]. Power System Technology, 2013, 37(06): 1719-1725. This reference is based on historical meteorological data of single base tower, uses radial neural network to predict various meteorological factors around the line in the next hour, and then gives the predicted value of line current carrying capacity based on conductor heat balance equation; Reference [5]: Alberti R, Albizu I, Fernandez E, et al. Overhead line ampacity forecasting with a focus on safety [J]. IEEE Transactions on Power Delivery, 2022, 37(1): 329-337. This reference is based on local meteorological data, uses single-layer perceptron model to predict the current carrying capacity of overhead transmission line; Reference [6]: Gao Z, Hu S, Sun H, et al. Day-ahead dynamic thermal line rating forecasting and power transmission capacity calculation based on ForecastNet [J]. Electric Power Systems Research, 2023, 220: 109350. This paper uses the ForecastNet model to dynamically track the time-varying characteristics and influence of various meteorological factors, achieving hourly forecasting of the current carrying capacity of overhead transmission lines. It should be noted that the above study, using a single forecasting model, has limitations in generalization ability and is difficult to maintain stable performance under different conditions.

[0007] In summary, existing research indicates that the relevant prediction methods suffer from the following problems: 1) Single prediction models are structurally limited, making it difficult to fully explore the deep characteristics of changes in various meteorological factors, resulting in insufficient generalization ability; 2) Traditional prediction models are trained using only historical data, which cannot characterize and describe meteorological prediction trends, and have the inherent limitation that the longer the time step, the greater the error; 3) Currently, due to the difficulty for power grid companies to deploy monitoring devices on all towers, many lines lack continuous micro-meteorological data, resulting in insufficient training of traditional current-carrying capacity models, making it impossible to achieve accurate prediction of current-carrying capacity across the entire span of overhead transmission lines. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning. This method integrates the advantages of CNN local feature extraction, GRU medium- and long-term time series modeling, and attention mechanism to focus on key information. Furthermore, it introduces and combines numerical weather prediction and transfer learning ideas to improve the accuracy of current carrying capacity prediction for target overhead transmission lines under the condition of small sample data constraints.

[0009] The technical solution adopted in this invention is as follows: A method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning includes the following steps: S1: Acquire numerical weather prediction (NWP) data and historical actual meteorological data of overhead transmission lines, and divide the data of transmission lines into source domain data and target domain data; S2: Subtract the numerical weather prediction (NWP) data of the source domain from the historical actual meteorological data of the overhead transmission line to obtain the initial tower micro-meteorological data error sequence; S3: Pre-trained model: A source domain pole micro-meteorological data error model was established using a CNN-GRU-attention hybrid algorithm to correct numerical weather prediction (NWP) data. After training, the model parameters were frozen and the model was saved. S4: Fine-tuning the model: Unfreeze the fully connected layer of the source domain tower micro-meteorological data error model, retrain the fully connected layer with the target domain error data, fine-tune the parameters of the source domain tower micro-meteorological data error model, and finally obtain the predicted values ​​of each meteorological factor in the target domain.

[0010] S5: Predicted values ​​of current carrying capacity of overhead transmission lines based on the heat balance equation.

[0011] This enables accurate prediction of the carrying capacity of the target line under the constraint of small sample data.

[0012] In S1, the numerical weather prediction (NWP) data is as follows: Figure 7 The dashed line in the image shows the historical meteorological data for overhead transmission lines, as shown in the image. Figure 7 The solid line portion is shown in the image.

[0013] In S2, the initial tower micro-meteorological data error sequence is as follows: (1); In formula (1): This is a meteorological error sequence; This is the NWP forecast value; These are measured data for the pole.

[0014] S2 can reduce the defect that the longer the step size, the greater the error.

[0015] S3 includes the following steps: S3.1: Obtain actual meteorological data of the towers along the source region transmission line and spatially match them with the corresponding numerical weather prediction (NWP) data; select the nearest NWP grid point as the forecast input for that tower; such as Figure 8 As shown, the blue broken line represents a route, which is matched to the corresponding grid points on the map. A source domain training sample set is constructed. D s .

[0016] S3.2: Establish an error model for source domain pole micro-meteorological data using a CNN-GRU-attention hybrid algorithm; and randomly initialize the model parameters. θ s The CNN-GRU-attention hybrid algorithm structure is as follows: Figure 9 As shown.

[0017] S3.3: Perform single-point adjustments to numerical weather prediction (NWP) data using source domain data; S3.4: Set the hyperparameters for the model pre-training phase, including the number of samples selected for each training session. S s and model learning rate L rs ; S3.5: Begin training the source domain tower micro-meteorological data error model, starting from the source domain training sample set. D s Selected from S s One sample is used as model input. x and their corresponding actual meteorological observation values y Update parameters θ s ; S3.6: After the model training is complete, freeze the model parameters and save the model.

[0018] The purpose of S3.1 to S3.6 above is to obtain a stable load capacity prediction model based on abundant source domain data.

[0019] In S3.2, in the CNN-GRU-attention hybrid algorithm, the CNN layer handles the meteorological error sequence between the measured tower data and the numerical weather prediction (NWP) data. X ={ x 1 , x 2 , x3 , ..., x T} i Feature vector x 1 , x 2 , x 3 ,..., x T These represent the meteorological errors for each time period; the convolution operation is calculated as follows: (3); In formula (3): h i CNN The output sequence of the CNN layer is passed to the GRU layer to extract the local temporal features of the micro-meteorological error sequence; x i Meteorological error series X The i-th eigenvector; k The kernel size; w i For convolution kernel; b This is a bias term.

[0020] In S3.2, the GRU layer operation mechanism in the CNN-GRU-attention hybrid algorithm is as follows: (4); (5); (6); (7); In the above formula: r z , z t These are the reset door and the update door, respectively. This is the candidate hidden state; The hidden state is passed to the next time step; σ represents the sigmoid activation function; h t-1 This is the hidden state from the previous moment; x t Enter information for the current moment; tanh is the activation function that assigns the data to [-1, 1]. This indicates resetting the weight of the door; This indicates the weight of the updated gate; The weights representing the candidate states; This indicates the offset of the door being reset; This indicates that the bias term of the updated gate is being updated; Bias terms representing candidate states; In S3.2, in the CNN-GRU-attention hybrid algorithm, the relevance score calculated based on the hidden state of the network in the Attention layer is used to adaptively assign weights to features at different time steps, thereby forming a new weighted feature sequence. The calculation formula is as follows: First, calculate the importance score of the time step features of the micro-meteorological data: (8); In equation (8): Indicates the attention relevance at time step t. This represents the meteorological error sequence at time t. and These are the weight matrix and the bias term, respectively.

[0021] Then, the correlation of all micrometeorological data time steps is normalized to obtain their attention weights: (9); (10); In the above formula: Let be the attention distribution function; This is to perform a weighted sum of input information through an information selection mechanism.

[0022] Finally, the micro-weather prediction value is passed to the output layer. y .

[0023] (11); In equation (11): It is the sigmoid activation function; W c , b c These are the weight matrix and the bias term, respectively.

[0024] The purpose of this is to assign higher weight values ​​to factors that have a significant impact.

[0025] S3.3 includes the following steps: a: Calculate the error sequence between NWP data and tower measured data, as shown in equation (1); b: Input the error sequence into the source domain tower micrometeorological data error model to obtain the error prediction value. .

[0026] c: The obtained error prediction value Adding this to the numerical weather prediction (NWP) data yields the corrected micro-weather forecast value.

[0027] (2); In formula (2): This is the revised microweather forecast value.

[0028] In step S4, a traffic prediction model for a small amount of data in the target domain is obtained based on transfer learning theory.

[0029] The theory of transfer learning is as follows: Transfer learning uses knowledge gained from one task or dataset to improve the performance of a model on another related task or different dataset. Transfer learning is based on domain... D and tasks T Defined as shown in the following two equations: domain D From the feature space χ and P ( X Composed of, among which X = { x 1, x 2, x 3, ..., x n}∈χ, task T From tag space y and target prediction function f : X → Y composition.

[0030] In transfer learning, models are typically pre-trained in a data-rich source domain and then fine-tuned in a data-limited target domain, thereby leveraging the implicit knowledge in the source domain data to improve the model's performance in the target domain.

[0031] (12); (13); In the above formula, The label space is represented. This represents the prediction function.

[0032] By leveraging transfer learning to capture common model features across different lines, we can compensate for the limitations imposed by data scarcity on prediction accuracy.

[0033] S4 includes the following steps: S4.1: The fully connected layer of the source domain tower micro-meteorological data error model is used to select existing meteorological data of the target domain transmission lines and construct a target domain training sample set. D t ; The parameters of the pre-trained model θ s Divided into feature extraction layer parameters θf With fully connected layer parameters θ c ; S4.2: Set the number of samples selected for each training session during the model fine-tuning phase. S t and model learning rate L rt .

[0034] S4.3: Fixed Feature Extraction Layer Parameters θ f Only for fully connected layer parameters θ c Update the training sample set from the target domain. D t Selected from S t One sample is used as input to the source domain tower micrometeorological data error model; S4.4: Repeat steps S4.2 and S4.3 until the optimal model parameters for the target domain data are obtained.

[0035] In step S4, during model transfer, only the fully connected layers of the model are unfrozen for parameter fine-tuning. Compared to training the model from scratch, fine-tuning only updates some network parameters, reducing the number of training parameters. This effectively constrains the model's complexity and helps improve its generalization ability when the target domain sample size is limited.

[0036] In S5, the heat balance equation is: (14); In equation (14): I Current in the conductor (unit: A); R ( T c ) represents the conductor at average temperature T c The resistance at that time (unit: Ω); Q s Heat absorption by the conductor under sunlight (unit: W / m); Q c For heat dissipation by convection in the conductor (unit: W / m); Q r Heat dissipation by radiation from the conductor (unit: W / m); T c , T a , V and J s These are conductor temperature, ambient temperature, wind speed, and solar radiation intensity, respectively.

[0037] In S5, the formula for calculating the current carrying capacity of overhead transmission lines is: (15); In the formula: Indicates the line's current carrying capacity; Q c Heat dissipation through air convection; Heat dissipation through radiation from the conductor; This represents the AC resistance per unit length of the conductor at that temperature.

[0038] This invention discloses a method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning. The technical effects are as follows: 1) This invention combines convolutional neural networks, gated recurrent units and attention mechanisms to construct a CNN-GRU-Attention serial structure hybrid model for micro-meteorology of overhead transmission lines. It can effectively utilize CNN, GRU and AM to achieve functions such as accurate local feature of micro-meteorology, accurate modeling of medium and long-term micro-meteorology, and focusing of key micro-meteorology information, so as to improve the accuracy of micro-meteorology prediction for overhead transmission lines.

[0039] 2) This invention introduces NWP data based on meteorological physics models, which contains micro-meteorological data with future meteorological trends, and can effectively improve the multi-step prediction accuracy and generalization ability of micro-meteorological prediction models.

[0040] 3) This invention introduces the concept of transfer learning, which effectively solves the problem of accurate prediction of the carrying capacity of the target line under the condition of small sample limitation. Attached Figure Description

[0041] The present invention will be further described below with reference to the accompanying drawings and examples; Figure 1 This is a flowchart of a method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning.

[0042] Figure 2 This is the environmental temperature prediction result after correction by the numerical weather prediction model.

[0043] Figure 3 This is the wind speed prediction result after correction by the numerical weather prediction model.

[0044] Figure 4 This is the solar radiation prediction result after correction by the numerical weather prediction model.

[0045] Figure 5 This is the revised carrying capacity prediction result from the numerical weather prediction model.

[0046] Figure 6 It is the traffic prediction result of the target domain after transfer learning.

[0047] Figure 7 This is a graph showing the historical actual meteorological data of numerical weather prediction (NWP) and overhead transmission lines.

[0048] Figure 8 This is a diagram illustrating how a route is matched to the corresponding grid points on a map.

[0049] Figure 9 This is a schematic diagram of the CNN-GRU-attention hybrid algorithm structure.

[0050] Figure 10(a) shows the source domain data. Figure 10(b) shows the target domain data. Detailed Implementation

[0051] This invention proposes a method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning. First, it uses a nearby line with abundant tower micrometeorological observation data as the source domain, fusing its historical micrometeorological data with numerical weather prediction (NWP) data that includes future trends to reduce multi-step prediction errors. Then, it constructs a CNN-GRU-Attention micrometeorological prediction model for this nearby line. Based on this, it combines the limited micrometeorological observation data already acquired for the target line and introduces transfer learning to establish a mapping relationship between the micrometeorological prediction values ​​of the nearby line and the predicted values ​​of the target line, thereby proposing a current carrying capacity prediction model for the target line and achieving accurate prediction of its current carrying capacity. Compared to traditional data-driven machine learning methods that require large amounts of data, this invention integrates the advantages of CNN local feature extraction, GRU medium-to-long-term time series modeling, and attention mechanisms to focus on key information. Furthermore, it incorporates and combines numerical weather prediction and transfer learning concepts, improving the accuracy of current carrying capacity prediction for the target overhead transmission line under conditions of limited sample data.

[0052] like Figure 1 As shown, a method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning is presented, with the following specific steps: Step S1: Obtain numerical weather prediction (NWP) data and historical actual meteorological data of the overhead transmission lines from two adjacent transmission lines, and divide the data of each transmission line into source domain data and target domain data. For example, data from a certain day is shown in Figures 10(a) and 10(b).

[0053] Step S2: Subtract the numerical weather prediction (NWP) data of the source domain from the historical actual meteorological data of the overhead transmission line to obtain the initial tower micro-meteorological data error sequence.

[0054] Step S3, Pre-trained model: A source domain pole micro-meteorological data error model was established using a CNN-GRU-attention hybrid algorithm to correct numerical weather prediction (NWP) data. After training, the model parameters were frozen and the model was saved.

[0055] The pre-trained model includes the following steps: ① Obtain measured meteorological data of towers along the source domain transmission line, spatially match them with the corresponding NWP data, select the nearest NWP grid point as the forecast input for that tower, and construct the source domain training sample set. D s .

[0056] ② Establish a CNN-GRU-attention hybrid model and randomly initialize the model parameters. θ s。 The hyperparameters of the model pre-training stage are set by single-point adjustment of the NWP data using source domain tower data, including the number of samples selected for each training session. S s And randomly initialize the model parameters. θ s and model learning rate L rs .

[0057] ③ Begin model training, starting with the training sample set from the source domain. D s Selected from S s One sample is used as model input. x and their corresponding actual meteorological observation values y Update parameters θ s .

[0058] ④ After the model training is complete, freeze the model parameters and save the model.

[0059] Step S4, fine-tune the model: Unfreeze the fully connected layer of the source domain tower micro-meteorological data error model, retrain the fully connected layer with the target domain error data, fine-tune the parameters of the source domain tower micro-meteorological data error model, and finally obtain the predicted values ​​of each meteorological factor in the target domain.

[0060] Fine-tuning the model includes the following steps: ①The fully connected layer of the thawing model selects existing measured meteorological data of transmission line towers in the target domain to construct a training sample set for the target domain. D t . The pre-trained model parameters θ s Divided into feature extraction layer parameters θ f With fully connected layer parameters θ c .

[0061] ② Set the number of samples selected for each training session during the model fine-tuning phase. S t and model learning rate L rt .

[0062] ③ Fix the parameters of the feature extraction layer θ f Only for fully connected layer parameters θ c Update the dataset from the target domain training sample set. D t Selected from S t Each sample is used as input to the model.

[0063] ④ Repeat steps ② and ③ until the optimal model parameters for the target domain data are obtained.

[0064] Step S5: Solve for the predicted value of the current carrying capacity of the overhead transmission line based on the heat balance equation.

[0065] The heat balance equation is: ; In the formula: I The current in the conductor is (A). R ( T c ) represents the conductor at average temperature T c The resistance (Ω) at that time; Q s Heat absorption by the conductor under sunlight (W / m); Q c For heat dissipation through convection in the conductor (W / m); Q r Heat dissipation through radiation from the conductor (W / m); T c , T a , V and J s These are conductor temperature, ambient temperature, wind speed, and solar radiation intensity, respectively.

[0066] ①: Q r The expression for heat dissipation by radiation from the conductor is: ; In the formula, T c D is the conductor temperature (°C); D is the conductor diameter (mm). T aε is the ambient temperature (°C). ε is the radiation coefficient of the conductor surface, which depends on the aging of the conductor. The value is [0.23, 0.43] for a bright new conductor and [0.90, 0.95] for an old conductor. σ is the Stefan-Boltzmann constant, which has a value of 5.67 × 10⁻⁸ (W·m⁻²·K⁻⁴).

[0067] ②: Qs represents the heat absorbed by the conductor under sunlight, and its expression is: ; In the formula, D The diameter of the conductor (mm); α s The heat absorption coefficient of the conductor surface ranges from 0.23 to 0.91. J s This represents the intensity of solar radiation.

[0068] ③: Q c For heat dissipation through air convection, the expression is: Air convection cooling occurs because air near a hot conductor is heated, reducing its density and causing airflow that carries away heat. The calculation formula is as follows: ; In the formula, D The diameter of the conductor (mm); k f The thermal conductivity of air (W / (m·℃)); ρ f Air density (kg / m3); μ f The dynamic viscosity of air (Pa·s); k angle It is a wind direction factor, and is related to the angle between the overhead power line and the wind direction; V w ε is the wind speed (m / s); ε is the heat dissipation coefficient of the conductor.

[0069] ④: Temperature T c The AC resistance value R(Tc) of the conductor at that time: First, the DC resistance of the wire needs to be calculated: ; In the formula, R 20 The conductor temperature is 20℃. T c τ is the conductor temperature (°C); τ is the temperature coefficient of resistance.

[0070] Then, calculate the AC resistance: ; In the formula,R z This represents the DC resistance value of the conductor. k is the skin effect coefficient of the conductor, and its magnitude is related to the cross-sectional area of ​​the conductor.

[0071] Substituting equations ①-④ into the heat balance equation, the formula for calculating the current carrying capacity of overhead transmission lines can be derived as follows: ; Then, the predicted meteorological error data is added to the NWP data, and then substituted into the carrying capacity calculation formula to obtain the carrying capacity prediction result.

[0072] Verification example: Simulation calculations were performed using a commonly used LGJ240 / 40 type overhead transmission line in a certain city's power grid. Specific line parameters are shown in Table 1. The calculation conditions for the static current carrying capacity were set as follows: ambient temperature 25℃, wind speed 0.5m / s, and solar radiation intensity 1000W / m².

[0073]

[0074] The prediction results are evaluated using mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R-squared, R²), calculated as follows: ; ; ; In the formula: n is the number of samples y i Indicates the actual value; This is a predicted value; This is the average value.

[0075] To verify the weather forecast correction effect of the proposed CNN-GRU-Attention model, its correction results are compared with those of GRU, LSTM, CNN-LSTM, and CNN-GRU models. The results are as follows: Figures 2-5 As shown, by Figures 2-5 As can be seen, after correction using the method of this invention, the consistency between the predicted ambient temperature, wind speed, solar radiation intensity and the actual situation of the tower is significantly improved, which strongly confirms that the prediction results obtained by this method can more accurately reflect the real meteorological conditions of the tower. The evaluation indicators of different prediction models are shown in Table 2.

[0076]

[0077] As shown in Table 2, compared with other models, the proposed model reduced the ERMSE of line current carrying capacity prediction by 69.92%, 69.31%, 7.16%, and 2.60%, respectively; the EMAE decreased by 73.47%, 77.88%, 10.31%, and 9%, respectively; and the R2 increased by 40.29%, 46.09%, 6.32%, and 2.09%, respectively.

[0078] To verify that introducing NWP data helps improve the accuracy of multi-step predictions, this invention sets different step sizes and different input data for comparison. The results are shown in Table 3 below.

[0079]

[0080] As can be seen from Table 3, the prediction error of the carrying capacity of the three sets of data increases with the increase of the prediction step size. However, the prediction accuracy of the NWP data set after correction is significantly higher than that of the other two sets. In the previous prediction, the ERMSE of the NWP data set after correction decreased by 34.91% and 51.43% respectively compared with the other two sets of data.

[0081] To verify the weather forecast correction effect of the proposed transfer learning method, we set a training set of one week in the target domain and compared the proposed model with the direct prediction model and the transfer direct prediction model. The prediction result for a certain day is as follows: Figure 6 As shown. From Figure 6 As can be seen, the accuracy of load capacity prediction is significantly improved after incorporating transfer learning. The evaluation metrics for load capacity prediction are shown in Table 4.

[0082]

[0083] As shown in Table 4, compared with the direct prediction model without transfer learning, the method of the present invention reduces ERMSE by 24.17%, EMAE by 25.44%, and R2 by 23.29%.

Claims

1. A method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning, characterized in that... Includes the following steps: S1: Acquire numerical weather prediction (NWP) data and historical actual meteorological data of overhead transmission lines, and divide the data of transmission lines into source domain data and target domain data; S2: Subtract the numerical weather prediction (NWP) data of the source domain from the historical actual meteorological data of the overhead transmission line to obtain the initial tower micro-meteorological data error sequence; S3: Use the CNN-GRU-attention hybrid algorithm to establish a source domain tower micro-meteorological data error model, correct the numerical weather prediction (NWP) data, freeze the model parameters after training, and save the model; S4: Unfreeze the fully connected layer of the source domain pole micro-meteorological data error model, retrain the fully connected layer with the target domain error data, fine-tune the parameters of the source domain pole micro-meteorological data error model, and finally obtain the predicted values ​​of each meteorological factor in the target domain. S5: Predicted values ​​of the current carrying capacity of overhead transmission lines based on the heat balance equation.

2. The method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning according to claim 1, characterized in that: In S2, the initial tower micro-meteorological data error sequence is as follows: （1）； In formula (1): This is a meteorological error sequence; This is the NWP forecast value; These are measured data for the tower.

3. The method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning according to claim 2, characterized in that: S3 includes the following steps: S3.1: Obtain actual meteorological data of the towers along the transmission line in the source domain and spatially match them with the corresponding numerical weather prediction (NWP) data; select the nearest NWP grid point as the forecast input for that tower; construct the source domain training sample set. D s ; S3.2: Establish an error model for source domain pole micro-meteorological data using a CNN-GRU-attention hybrid algorithm; and randomly initialize the model parameters. θ s ; S3.3: Perform single-point adjustments to numerical weather prediction (NWP) data using source domain data; S3.4: Set the hyperparameters for the model pre-training phase, including the number of samples selected for each training session. S s and model learning rate L rs ; S3.5: Begin training the source domain tower micro-meteorological data error model, starting from the source domain training sample set. D s Selected from S s One sample is used as model input. x and their corresponding actual meteorological observation values y Update parameters θ s ; S3.6: After the model training is complete, freeze the model parameters and save the model.

4. The method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning according to claim 3, characterized in that: In S3.2, in the CNN-GRU-attention hybrid algorithm, the CNN layer handles the meteorological error sequence between the measured tower data and the numerical weather prediction (NWP) data. X ={ x 1 , x 2 , x 3 , ..., x T } i Feature vector x 1 , x 2 , x 3 , ..., x T These represent the meteorological errors for each time period; the convolution operation is calculated as follows: （3）； In formula (3): h i CNN The output sequence of the CNN layer is passed to the GRU layer to extract the local temporal features of the micro-meteorological error sequence; x i Meteorological error series X The i-th eigenvector; i>k is the kernel size; w i For convolution kernel; b This is a bias term.

5. The method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning according to claim 4, characterized in that: In S3.2, the GRU layer operation mechanism in the CNN-GRU-attention hybrid algorithm is as follows: （4）； （5）； （6）； （7）； In the above formula: r z , z t These are the reset door and the update door, respectively. This is the candidate hidden state; The hidden state is passed to the next time step; σ represents the sigmoid activation function; h t-1 This is the hidden state from the previous moment; x t Enter information for the current moment; tanh is the activation function that assigns the data to [-1, 1]. This indicates resetting the weight of the door; This indicates the weight of the updated gate; The weights representing the candidate states; This indicates the offset of the door being reset; This indicates that the bias term of the updated gate is being updated; The bias term represents the candidate state.

6. The method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning according to claim 5, characterized in that: In S3.2, in the CNN-GRU-attention hybrid algorithm, the relevance score calculated based on the hidden state of the network in the Attention layer is used to adaptively assign weights to features at different time steps, thereby forming a new weighted feature sequence, specifically including: First, calculate the importance score of the time step features of the micro-meteorological data: （8）； In equation (8): Indicates the attention relevance at time step t. This represents the meteorological error sequence at time t. and These are the weight matrix and the bias term, respectively. Then, the correlation of all micrometeorological data time steps is normalized to obtain their attention weights: （9）； （10）； In the above formula: Let be the attention distribution function; To perform a weighted sum of input information through an information selection mechanism; Finally, the micro-weather prediction value is passed to the output layer. y ; （11）； In equation (11): It is the sigmoid activation function; W c , b c These are the weight matrix and the bias term, respectively.

7. The method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning according to claim 6, characterized in that: S3.3 includes the following steps: a: Calculate the error sequence between NWP data and tower measured data, as shown in equation (1); b: Input the error sequence into the source domain tower micrometeorological data error model to obtain the error prediction value. ; c: The obtained error prediction value Adding this to the numerical weather prediction (NWP) data yields the corrected micro-weather forecast value; （2）； In formula (2): This is the revised microweather forecast value.

8. The method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning according to claim 7, characterized in that: In step S4, a traffic prediction model for a limited amount of data in the target domain is obtained based on transfer learning theory; specifically as follows: Transfer learning is a domain D and tasks T Defined as follows: domain D From the feature space χ and P ( X Composed of, among which X = { x 1, x 2, x 3, ..., x n }∈χ, task T From tag space y and target prediction function f : X → Y composition; In transfer learning, the model is pre-trained in a data-rich source domain and then fine-tuned in a data-limited target domain, thereby leveraging the implicit knowledge in the source domain data to improve the model's performance in the target domain. （12）； （13）； In the above formula, The label space is represented. This represents the prediction function.

9. The method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning according to claim 8, characterized in that: S4 includes the following steps: S4.1: The fully connected layer of the source domain tower micro-meteorological data error model is used to select existing meteorological data of the target domain transmission lines and construct a target domain training sample set. D t ; The parameters of the pre-trained model θ s Divided into feature extraction layer parameters θ f With fully connected layer parameters θ c ; S4.2: Set the number of samples selected for each training session during the model fine-tuning phase. S t and model learning rate L rt ; S4.3: Fixed feature extraction layer parameters θ f Only for fully connected layer parameters θ c Update the training sample set from the target domain. D t Selected from S t One sample is used as input to the source domain tower micrometeorological data error model; S4.4: Repeat steps S4.2 and S4.3 until the optimal model parameters for the target domain data are obtained.

10. The method for predicting the current carrying capacity of overhead transmission lines based on numerical weather prediction model correction and transfer learning according to claim 9, characterized in that: In S5, the heat balance equation is: （14）； In equation (14): I Current in the conductor (unit: A); R ( T c ) represents the conductor at average temperature T c The resistance at that time (unit: Ω); Q s Heat absorption by the conductor under sunlight (unit: W / m); Q c For heat dissipation by convection in the conductor (unit: W / m); Q r Heat dissipation by radiation from the conductor (unit: W / m); T c , T a , V and J s These are conductor temperature, ambient temperature, wind speed, and solar radiation intensity, respectively. The formula for calculating the current carrying capacity of overhead transmission lines is: （15）； In the formula: Indicates the line's current carrying capacity; Q c Heat dissipation through air convection; Heat dissipation through radiation from the conductor; This represents the AC resistance per unit length of the conductor at that temperature.

Images