Unmanned aerial vehicle automatic airport operation power prediction method

Abstract

The invention provides an unmanned aerial vehicle automatic airport operation power prediction method, and relates to the field of machine learning. The unmanned aerial vehicle automatic airport operation power prediction process provided by the invention comprises the steps of constructing an unmanned aerial vehicle automatic airport operation power prediction data set, detecting and replacing abnormal values in the data set by using a Z-Score method and a sliding window mean value replacement method, and then realizing factorization of data features through a multi-layer perceptron. The method comprises the following steps: performing factorization on features to remove noise and redundancy therein, mapping the factorized features into high-dimensional features by using a lookup table, further extracting deep-layer features by combining a self-attention mechanism and a layer normalization method, then performing dimension reduction again by using a linear mapping method to remove noise, and finally obtaining a feature vector containing overall historical information by using a long short-term memory (LSTM) network. According to the method, noise and redundancy in the data can be effectively removed, and the generalization ability and accuracy of prediction of the method are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of machine learning, and in particular relates to a method for predicting the operating power of an unmanned aerial vehicle automated airport. Background Art

[0002] The operating power of a drone automated airport refers to the power output required to support drone operations such as takeoff, landing, charging, communication and data processing in a drone automated airport. Due to the efficient operation of the automation system, this power demand usually changes dynamically. If the power is overloaded or unstable, it may cause equipment overheating, system failure or energy waste. Therefore, predicting the operating power of a drone automated airport can help operators optimize energy allocation, improve system reliability, and reduce operational interruptions or economic losses caused by power problems, thereby achieving more efficient and sustainable automated airport operations.

[0003] Research on power prediction for UAV automated airport operations mainly relies on traditional statistical methods and machine learning models. Traditional statistical methods perform well in dealing with linear relationships, but have obvious limitations when dealing with nonlinear and high-dimensional feature data. In contrast, machine learning methods, such as support vector machines and random forests, have significant advantages in modeling nonlinear relationships. They can not only capture nonlinear characteristics, but also achieve high-precision predictions for long time series data, thereby providing reliable technical support for energy management and operation optimization of UAV automated airports.

[0004] The data sources for the prediction of the operating power of UAV automated airports are diverse and complex, and are significantly affected by noise interference. This brings many challenges to the application of machine learning technology in data processing and modeling. Therefore, it is particularly important to apply denoising technology in the prediction of the operating power of UAV automated airports. By effectively removing the noise in the original data, not only can the data quality be significantly improved, but also the generalization ability and robustness of the model can be improved, thereby improving the applicability and reliability of the prediction model in complex practical scenarios. Summary of the invention

[0005] The present invention provides a method for predicting the operating power of an unmanned aerial vehicle automated airport, which aims to detect and replace outliers in data by combining a Z-Score method and a sliding window mean replacement method, thereby improving data quality, and using a factor decomposition module to remove noise and redundant features in the data. Subsequently, a lookup table is used to implement high-dimensional feature embedding to enhance feature expression capabilities, and then a self-attention mechanism is used to extract deep key features, and dimensionality reduction operations are combined to further remove noise and improve the generalization ability of the model. On this basis, a long short-term memory network is used to capture the overall historical information of the time series, and finally a fully connected neural network is used to accurately predict the operating power of an unmanned aerial vehicle automated airport.

[0006] The technical method adopted by the present invention to achieve the above-mentioned purpose specifically comprises the following steps: S1. Collect relevant monitoring data of drone automated airports, including wind speed, wind direction, number of drones in operation, drone task scheduling, charging equipment power, charging time of charging equipment, and health status of charging equipment. Use the Z-Score method to detect outliers in the collected data and replace the outliers with the average value in the sliding window, so as to construct a high-quality drone automated airport operation power prediction data set, which provides a solid foundation for subsequent analysis and modeling; S2. Build a factor decomposition module to factorize features through a multi-layer perceptron to remove noise and redundancy in multivariate time series data, thereby reducing computational complexity and extracting representative and effective features; S3, feature embedding is achieved through lookup tables, which maps input data into a high-dimensional feature space to enhance feature expression capabilities; S4. Build a deep feature extraction module, use the self-attention mechanism to effectively extract the deep features of time series data, capture its long-term dependencies, and combine the layer normalization method to keep the deep features stable; S5, perform dimensionality reduction operation, use linear mapping method to complete dimensionality reduction for high-dimensional data of each time step, so as to remove noise components and retain the most informative key features; S6. Build a long short-term memory network LSTM feature extraction module, process long time series data through the long short-term memory network LSTM, so that the output of the last time step contains the overall historical information of the time series; S7. Construct a fully connected neural network and calculate the final feature vector to obtain the final UAV automated airport operation power prediction value.

[0007] Preferably, in S1, relevant monitoring data of the drone automated airport is collected, including wind speed, wind direction, number of drones in operation, drone task scheduling, charging device power, charging device charging time, and health status of the charging device, wherein the i-th type of data sequence collected is recorded as Where m is the length of time for collecting data, is the monitoring value of the i-th data at the j-th time step. The Z-Score method is used to detect abnormal values ​​in each data. Taking the i-th data as an example, the specific calculation process is as follows: In the formula For data point x i，j The standard deviation multiple of the deviation from the mean, where a positive value indicates that the data point is above the mean and a negative value indicates that it is below the mean, μ i is the mean of the ith data distribution, σi is the standard deviation of the ith data distribution, if Z i，j If the absolute value of x exceeds a certain threshold, the data x i，j The outliers are regarded as outliers, and a sliding window of size S is constructed with the outliers as the center. The mean of other data points in the window is calculated and the outliers are replaced. When the window exceeds the boundary, the data points outside the boundary are ignored. Finally, a high-quality UAV automated airport operation power prediction dataset is constructed through the above steps, laying the foundation for subsequent analysis and prediction.

[0008] Preferably, multi-dimensional drone automated airport monitoring data is collected, and the Z-Score method is used to detect outliers. The sliding window mean replacement method is used to process outliers. This can ensure the quality and consistency of the data and effectively reduce the impact of noise on subsequent analysis. Compared with directly deleting outliers, this method can retain more useful information, thereby improving the accuracy and robustness of model predictions.

[0009] Preferably, in S2, the time series data of the input UAV automated airport operation power prediction is recorded as Where T is the length of the input time series, n=7 is the number of data types collected by S1, is all types of data collected at time step t, where is the monitoring value obtained by the i-th data at the t∈[1, T]th time step. A multilayer perceptron is used on the time series data Y to map the original data to hidden layer features and map the hidden layer features back to the output space, thereby realizing factor decomposition. The specific calculation process is as follows: In the formula is the feature matrix after factorization, is the eigenvector after factorization at the tth time step, Y is the original input time series data matrix, is the learnable weight parameter, is the bias term, and σ(·) is the ReLU activation function.

[0010] Preferably, the core idea of ​​the factor decomposition process is to use a multi-layer perceptron to extract the potential features of time series data, map it from the input space to the latent space and then map it back to the original space through a nonlinear transformation, thereby removing noise and retaining key information. Through the above steps, the factor decomposition of the UAV automated airport operation power prediction data can be effectively realized to achieve the purpose of denoising, better adapt to the changes in data features in complex scenarios, provide high-quality input data for subsequent model construction, and improve the accuracy and stability of the prediction.

[0011] Preferably, in S3, in order to further extract features and vectorize the time series feature matrix Y′ after factor decomposition, a lookup table function is used to map each value to a high-dimensional feature space. The specific calculation process is as follows: V=Emb(Y′); In the formula is the new feature matrix generated by the lookup table function of Y′, where q=n×d is the feature dimension of each time step feature after being mapped by the lookup table, n is the number of data types of the original data, and d is the feature vector dimension after the lookup table maps each value. is the feature vector composed of all types of data at the tth time step, and Emb(·) is a lookup table function, which maps each value of the input matrix Y′ to a d-dimensional feature vector, thereby enhancing the feature representation ability of the input data.

[0012] Preferably, the introduction of the lookup table function effectively utilizes the embedding expression ability of discrete values ​​and converts each numerical value of the data into a high-dimensional feature vector. Compared with directly using numerical features, this method can better capture the potential semantic associations between data and enhance the robustness and generalization ability of the model to diverse features.

[0013] Preferably, in S4, a deep feature extraction module is constructed, firstly, the dependency between features of each time step is calculated using the self-attention mechanism, the features are weighted summed by the attention weights, and then the results are applied layer normalization. The specific calculation process is as follows: L = layerNorm(L′); Where s t，i is the correlation score of the feature vector between the t-th time step and the i-th time step, v i The transpose of t，i To t，i The normalized attention weights are is the weighted feature vector based on the attention weight, is the feature matrix calculated by the self-attention mechanism, layerNorm(·) is the layer normalization operation, is the final deep feature matrix obtained after applying layer normalization.

[0014] Preferably, the self-attention mechanism can dynamically adjust the weights between features to highlight the contribution of key data, while layer normalization further improves the model's adaptability to changes in input data. Through the combination of self-attention mechanism and layer normalization, it can capture complex dependencies in time series data while alleviating the impact of changes in feature distribution on subsequent modeling.

[0015] Preferably, in S5, a dimensionality reduction operation is performed on the final deep feature matrix L, and a linear mapping method is used to map the original high-dimensional features to a lower dimension to remove noise components and retain the most informative key features. The specific calculation process is as follows: R = ReLU(LWT+b); In the formula is the data feature matrix after dimensionality reduction, where k = q / 2 is the dimension after dimensionality reduction, is the eigenvector after dimensionality reduction of the eigenvector at time step t, is the learnable weight parameter, is the bias term, and ReLU(·) is the ReLU activation function.

[0016] Preferably, the dimensionality reduction of the deep feature matrix is ​​achieved through linear mapping, and the noise components are effectively removed by using learnable weight parameters and bias terms, while retaining the most informative key features in the data. Compared with directly retaining the original high-dimensional features, the dimensionality reduction operation can reduce the computational complexity, improve the training efficiency of the model, and avoid the redundancy problems that may exist in high-dimensional data.

[0017] Preferably, in S6, a long short-term memory network LSTM feature extraction module is constructed, and the long short-term memory network LSTM is used to capture long-term dependencies to obtain a final feature vector containing the entire historical information. The specific calculation process is as follows: i t =σ(r t W i，r +h t-1 W i，h +b i ); f t =σ(r t W f，r +h t-1 W f，h +b f ); o t =σ(r t W o，r +h t-1 W o，h +b o ); h t =o t ⊙tanh(c t ); Where W i，r , W i，h , W f，r , W f，h , Wo，r , W o，h , W c，r , is the learnable weight parameter, b i 、b f 、b o , is the bias term, σ(·) is the Sigmoid function, i t 、f t , They are the input gate, forget gate and output gate of the t-th time step of the long short-term memory network LSTM, is the candidate memory unit at the tth time step, is the memory unit at the tth time step, is the hidden state at the t-th time step, ⊙ is the element-wise multiplication, and tanh(·) is the tanh function.

[0018] Preferably, the input gate of the long short-term memory network LSTM selectively writes the new information of the current time step into the memory unit, the forget gate retains key information and ignores useless historical information through a screening mechanism, and the output gate further determines how the information in the memory unit is transmitted to the hidden state. Through these three gating mechanisms, the update and fusion of input features, historical memory and output features are dynamically controlled to effectively capture the long-term and short-term dependencies in the time series data, so that the output of the last time step contains the overall historical information of the time series, providing a basis for subsequent analysis and prediction.

[0019] Preferably, in S7, the output of the last time step of the long short-term memory network LSTM is taken as the final feature vector, and the feature vector is passed through a fully connected neural network to calculate the final UAV automated airport operation power prediction value. The specific calculation process is as follows: P = FCNN(O T ); Where FCNN(·) is a fully connected neural network, O T is the output of the last time step of the long short-term memory network LSTM, Power predictions for eventual drone automation at airports.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention detects and replaces outliers in the data by combining the Z-Score method and the sliding window mean replacement method, thereby effectively improving the data quality; in subsequent feature processing, the factor decomposition technology is used to remove data redundancy and noise, and the data is mapped to a high-dimensional space through a lookup table to enhance the feature expression capability; the self-attention mechanism is used to capture the deep features of the data, and the linear mapping further removes redundant information and strengthens the key feature representation; in addition, a long short-term memory network is used to obtain a unique feature vector containing the key historical information of the entire time series, and finally the unique feature vector is used to obtain a prediction result through a fully connected neural network, thereby significantly improving the generalization ability and accuracy of the prediction of the operating power of drone automated airports. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A step diagram of a method for predicting power of UAV automated airport operations.

[0022] Figure 2 It is the factor decomposition module diagram.

[0023] Figure 3 This is the deep feature extraction module diagram.

[0024] Figure 4 This is the LSTM feature extraction module diagram for the long short-term memory network.

[0025] Figure 5 Power prediction effect diagram for automated airport operation of drones. DETAILED DESCRIPTION

[0026] The present invention proposes a method for predicting the operating power of an automated airport of an unmanned aerial vehicle. The method detects and replaces outliers through a Z-Score method combined with a sliding window mean replacement method, repairs the noise data, and then reduces redundancy and extracts key features through a factor decomposition module. The data is then mapped to a high-dimensional space through a lookup table to enhance the feature expression capability. The self-attention mechanism and layer normalization technology are combined to further extract the deep features of the time series to capture its complex patterns. Then, linear mapping is used to remove noise in the high-dimensional features, and a long short-term memory network LSTM is used to capture the overall information of the time series, so as to obtain the prediction result through a fully connected neural network calculation. The technical solution in the embodiment of the present invention is described in detail and completely below, specifically including the following steps: Figure 1 shown.

[0027] S1. Collect 92 days of monitoring data of seven types of drone automated airports, including wind speed, wind direction, number of drones in operation, drone task scheduling, charging equipment power, charging time of charging equipment, and health status of charging equipment. Use the Z-Score method to detect outliers in the collected data, and then replace the outliers with the average value in the sliding window to construct a high-quality drone automated airport operation power prediction dataset. The dataset is divided into a training set and a validation set in a ratio of 7 to 3, and the first 7 days of data for each day are used as historical data for method learning.

[0028] Furthermore, in S1, seven types of monitoring data are collected every day at the drone automated airport, including wind speed, wind direction, number of drones in operation, drone task scheduling, charging equipment power, charging time of charging equipment, and health status of charging equipment. The Z-Score method is used to detect abnormal values ​​in each type of data. Taking the i-th type of data among the seven types of data as an example, the i-th type of data sequence is recorded as Where m = 92 is the total number of days for collecting data, is the monitoring value of the i-th data at the j-th time step. The specific calculation process is as follows: In the formula For data point x i，j The number of standard deviations from the mean, μ i is the mean of the ith data distribution, σ i is the standard deviation of the ith data distribution, Z i，j A positive value indicates that the data point is above the mean, and a negative value indicates that it is below the mean. i，j |≥3, then the data x i，j Considered as an outlier, x i，j Build a sliding window of size 3 for the center, calculate the mean of other data points in the window and replace x i，j ,If the window exceeds the boundary, the data points outside the boundary are ignored.,Finally, the outliers in the UAV automated airport operation power prediction dataset are removed through the above steps, and then the dataset is divided into a training set and a validation set in a ratio of 7 to 3, and the first 7 days of data for each day are used as the input for method learning.

[0029] S2. Construct a factor decomposition module, map the data from the input space to the latent space through a multi-layer perceptron, and then map it back from the latent space to the original space, to achieve feature factorization, remove redundancy in multivariate time series data, and extract representative and effective features.

[0030] Furthermore, in S2, the factor decomposition module is as follows: Figure 2As shown, the input T = 7 size of the UAV automated airport operation power prediction data of the continuous time series is recorded as Where n=7 is the number of data types collected by S1. is all types of data collected at the t∈[1,7] time step, where is the monitoring value obtained by the i-th data at the t-th time step. The multi-layer perceptron is used to factorize the time series data Y. The specific calculation process is as follows: In the formula is the feature matrix after factorization, is the eigenvector after factorization at the tth time step, Y is the original input time series data matrix, and are learnable weight parameters and bias items, which are used to map raw data to hidden layer features. and are learnable weight parameters and bias terms, respectively, used to map the hidden layer features back to the output space. σ(·) is the ReLU activation function, which performs nonlinear transformation on the input to enhance the model's expressiveness.

[0031] S3. Map the input data to a high-dimensional feature space through a lookup table to achieve feature embedding to enhance feature expression capabilities.

[0032] Furthermore, in S3, a lookup table function is used to map each value of the time series feature matrix Y′ to a high-dimensional feature space for feature extraction and vectorization processing. The specific calculation process is as follows: V = Emb(Y′); In the formula is the new feature matrix generated by the lookup table function of Y′, and Emb(·) is the lookup table function, which maps each element value of the input matrix Y′ to a 32-dimensional feature vector. Therefore, 224=7×32 is the number of data types of the original data multiplied by the feature vector dimension of each value after the lookup table mapping. It is the feature vector composed of all types of data at the tth time step.

[0033] S4. Build a deep feature extraction module, pass the new feature matrix generated by the lookup table function through the self-attention mechanism to extract the deep features of the time series data, capture its long-term dependencies, and then further process it using the layer normalization method to keep the features stable.

[0034] Furthermore, in S4, the deep feature extraction module is as follows: Figure 3As shown in the figure, the self-attention mechanism is first used to calculate the correlation scores between all time step features, and the correlation scores are normalized to obtain the attention weights. At this time, the features can be weighted summed according to the attention weights and then layer normalization can be performed. The specific calculation process is as follows: L = layerNorm(L′); Where s t，i is the correlation score of the feature vectors between the t-th time step and the i-th time step, v i The transpose of t，i For t，i The attention weight after normalization operation, is the result of weighted sum of other feature vectors based on the corresponding attention weight at the tth time step, and the results of all time steps are combined to obtain the feature matrix is the final deep feature matrix obtained after layer normalization, and LayerNorm(·) is the layer normalization operation.

[0035] S5. After the deep feature extraction module, the high-dimensional data at each time step contains a large amount of information, which is easily mixed with noise components. Therefore, the linear mapping method is used to reduce the dimensionality of the high-dimensional data to remove the noise components and retain the most informative key features.

[0036] Furthermore, in S5, a linear mapping method is used to perform a dimensionality reduction operation, mapping the deep feature matrix L from the original high-dimensional features to a lower dimension to remove noise components and retain the most critical features. The specific calculation process is as follows: R; In the formula is the data feature matrix after dimensionality reduction, where k = 224 / 2 = 112 is the dimension after dimensionality reduction, is the data feature vector of the tth time step after dimensionality reduction, is the learnable weight parameter, is the bias term, and ReLU(·) is the ReLU activation function.

[0037] S6. Construct a long short-term memory network LSTM feature extraction module, process the reduced-dimensional data feature matrix through the long short-term memory network LSTM, and integrate all key historical information into the output vector of the last time step, so that the output vector of the last time step can represent the information of the entire time series data.

[0038] Furthermore, in S6, the long short-term memory network LSTM feature extraction module is as follows: Figure 4As shown in the figure, the long short-term memory network LSTM is used to capture long-term dependencies and obtain the final feature vector containing the entire key historical information. The specific calculation process is as follows: i t =σ(r t W i，r +h t-1 W i，h +b i ); f t =σ(r t W f，r +h t-1 W f，h +b f ); o t =σ(r t W o，r +h t-1 W o，h +b o ); h t =o t ⊙tanh(c t ); Where i t 、f t , are the input gate, forget gate and output gate of the t-th time step of the long short-term memory network LSTM, respectively. i，r , W i，h , W f，r , W f，h , W o，r , W o，h , W c，r , is the learnable weight parameter, b i 、b f 、b o , is the bias term, σ(·) is the Sigmoid function, is the candidate memory unit at the tth time step, is the memory unit at the tth time step, is the hidden state at the t-th time step, ⊙ is the element-wise multiplication, and tanh(·) is the tanh function.

[0039] S7. Construct a fully connected neural network, take the output vector of the last time step of the long short-term memory network LSTM as input, and calculate the final predicted value of the UAV automated airport operation power.

[0040] Furthermore, in S7, the output vector o of the last time step of the long short-term memory network LSTM is 7 As the input of the fully connected neural network, the feature vector is calculated to obtain the predicted value of the UAV automated airport operation power. The specific calculation process is as follows: P = FCNN(o 7 ); In the formula is the final UAV automated airport operation power prediction value, FCNN(·) is a fully connected neural network, o 7 is the output vector of the last time step of the long short-term memory network LSTM.

[0041] Furthermore, the language used in this method is Python 3.8, and the model of this method is implemented using the PyTorch framework. The NVIDIA RTX 3090 GPU is used for training in the CUDA 12 environment. The batch size during training is 8 and the learning rate is 1×10 -4 , the optimizer is AdamW, and the loss function is mean square error MSE.

[0042] Furthermore, the method predicts the effect as Figure 5 As shown, the ordinate is the UAV automated airport operating power value (watt), the abscissa is time (day), the gray dotted line is the actual value of the UAV automated airport operating power, and the black straight line is the predicted value of the UAV automated airport operating power. Figure 5 Observation and analysis show that the predicted value of the UAV automated airport operating power obtained by this method shows a good fit with the true value, which proves the effectiveness of this method in the scenario of UAV automated airport operating power prediction.

[0043] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for predicting the operating power of an unmanned aerial vehicle automated airport, characterized in that: The following steps are involved: S1. Collect 7 types of drone automated airport monitoring data, including wind speed, wind direction, number of drones in operation, drone task scheduling, charging equipment power, charging equipment charging time, and charging equipment health status. Use the Z-Score method to detect outliers in each data, then build a sliding window of size S=3 with the outlier as the center, calculate the average value of other data points in the window to replace the outlier, and thus build a high-quality drone automated airport operation power prediction data set; S2. Build a factor decomposition module, map the data from the input space to a higher-dimensional implicit space through a multi-layer perceptron and then map it back to the original space, realize channel feature factorization, remove redundancy in time series data, and thus extract more representative features; S3, feature embedding is achieved through lookup tables, which maps input data into a high-dimensional feature space to enhance feature expression capabilities; S4. Build a deep feature extraction module. First, use the self-attention mechanism to effectively extract the deep features of time series data and capture its long-term dependencies. Then, perform layer normalization to keep the features of each layer stable. S5, perform dimensionality reduction operation. After the deep feature extraction module, the high-dimensional information of each time step is easily mixed with noise components. Therefore, the linear mapping method is used to reduce its dimension to remove the noise components and retain the most informative key features. S6. Build a long short-term memory network LSTM feature extraction module to process long time series data through the long short-term memory network LSTM. At this time, the output feature vector of the last time step contains the overall historical information of the time series, providing support for subsequent predictions; S7. Construct a fully connected neural network and calculate the output vector of the last time step of the long short-term memory network LSTM to obtain the final UAV automated airport operation power prediction value.

2. The method for predicting the operating power of an unmanned aerial vehicle automated airport according to claim 1, characterized in that: S1 collects the monitoring values ​​of wind speed, wind direction, number of operating drones, drone task scheduling, charging equipment power, charging time of charging equipment, and health status of charging equipment in the drone automated airport every day. Taking the i-th type of data as an example, after collecting m days, the monitoring values ​​are arranged in chronological order to form time series data and recorded as is the monitoring value of the i-th data at the j-th time step. The Z-Score method is used to detect abnormal values ​​in each data. Taking the i-th data as an example, the specific calculation process is as follows: In the formula For data point x i，j The number of standard deviations from the mean. Positive values ​​indicate that x i，j Above the mean, negative values ​​indicate below the mean, μ i is the mean of the ith data distribution, σ i is the standard deviation of the ith data distribution, if Z i，j If the absolute value of x exceeds the specified threshold 3, the data x i，j is considered as an outlier, and x i，j Build a sliding window of size S as the center and calculate the mean of other data points in the window to convert x i，j Replace it with the mean. When the sliding window exceeds the boundary, ignore the data points outside the boundary. Finally, after replacing all outliers, a high-quality drone automated airport operation power prediction dataset can be obtained, laying the foundation for subsequent analysis and modeling.

3. The method for predicting the operating power of an unmanned aerial vehicle automated airport according to claim 2, characterized in that: In S2, in the constructed data set, a sliding window is constructed with T as the time series length, and the data in the sliding window is used as the time series data input by the method, denoted as n=7 is the number of data types collected by S1, including wind speed, wind direction, number of UAVs in operation, UAV task scheduling, charging equipment power, charging equipment charging time, and health status of charging equipment. are all types of data values ​​collected at time step t, where is the monitoring value obtained by the i-th data at the t∈[1, T]th time step. The time series data Y is mapped to the hidden layer features using a multilayer perceptron and then mapped back to the output space to achieve factor decomposition. The specific calculation process is as follows: In the formula is the feature matrix after factorization, is the eigenvector after factorization at the tth time step, is the learnable weight parameter, is the bias term, Y is the time series data matrix of the original input, and σ(·) is the ReLU activation function.

4. The method for predicting the operating power of an unmanned aerial vehicle automated airport according to claim 3 is characterized in that: S4, after feature embedding through the lookup table function of S3, obtains the feature matrix Among them, q = n × d is the feature dimension of each time step feature after being mapped by the lookup table, n is the number of data types of the original data, d is the feature vector dimension after the lookup table maps each value, and the feature matrix V is input into the self-attention mechanism. First, the correlation scores between all time step features are calculated, and the correlation scores are normalized to obtain the attention weights. Finally, the features are weighted and summed according to the attention weights. The specific calculation process is as follows: Where s t，i is the correlation score of the feature vector between the t-th time step and the i-th time step, v i The transpose of t，i To t，i Normalized attention weights, is the weighted feature vector based on the attention weight. Then the feature vector after the self-attention mechanism is layer-normalized. The specific calculation process is as follows: L = layerNorm(L') In the formula is the final deep feature matrix obtained after applying layer normalization, is the feature matrix calculated by the self-attention mechanism, and layerNorm(·) is the layer normalization operation.

5. A method for predicting the operating power of an unmanned aerial vehicle automated airport according to claim 4, characterized in that: S5, using a linear mapping method to perform a dimensionality reduction operation on the final deep feature matrix L, by mapping the original high-dimensional features to low-dimensional features, removing the noise components in the features and retaining the most critical information. The specific calculation process is as follows: In the formula is the data feature matrix after dimensionality reduction through linear mapping, where k = q / 2 is the dimension after dimensionality reduction, is the feature vector after dimensionality reduction of the data feature vector at time step t, is the learnable weight parameter, is the bias term, and ReLU(·) is the ReLU activation function.

6. A method for predicting the operating power of an unmanned aerial vehicle automated airport according to claim 5, characterized in that: S6 constructs a long short-term memory network LSTM feature extraction module, uses the long short-term memory network LSTM to capture long-term dependencies, so that the output vector of the last time step contains the overall historical information of the time series. The specific calculation process is as follows: i t =σ(r t W i，r +h t-1 W i，h +b i ); f t =σ(r t W f，r +h t-1 W f，h +b f ); o t =σ(r t W o，r +h t-1 W o，n +b o ); h t =o t ☉tanh(c t ); In the formula They are the input gate, forget gate and output gate of the t-th time step of the long short-term memory network LSTM, is the learnable weight parameter, is the bias term, σ(·) is the Sigmoid function, is the candidate memory unit at the tth time step, is the memory unit at the tth time step, is the hidden state at the t-th time step, ⊙ is the element-wise multiplication, and tanh(·) is the tanh function.

7. The method for predicting the operating power of an unmanned aerial vehicle automated airport according to claim 1, characterized in that: In order to predict the operating power of drone automated airports, relevant monitoring data of drone automated airports were collected, including wind speed, wind direction, number of drones in operation, drone task scheduling, charging equipment power, charging time of charging equipment and health status of charging equipment. The Z-Score method combined with the sliding window mean replacement method was used to detect and replace outliers for each data. Then, the data set after processing the outliers was divided into training set and validation set in a ratio of 7 to 3 for training and validation of the method, which laid a reliable data foundation for the subsequent analysis and prediction of the operating power of drone automated airports by this method.

Images