Feature engineering method and system suitable for load prediction feature re-calibration

Abstract

The invention discloses a feature engineering method and system suitable for load prediction feature recalibration, and the method comprises the steps: obtaining historical load data and meteorological original data of a target region, and unifying the time span of the historical load data and the meteorological original data; splicing and anomaly detection are carried out on the historical load data and the meteorological original data after the time span is unified, Gaussian white noise disturbance is manually added after detection, and an initial sample set is constructed; basic features based on the initial sample set are constructed, feature combination is carried out on the basic features, feature screening is carried out by using a puller method and a filter method, and a feature matrix is generated; inputting the feature matrix into an SE-Net model, and performing feature matrix compression and remapping to obtain a feature sample set; splicing the initial sample set and the feature sample set, constructing a training data set, and performing fitting, verification and prediction by using an LSTM model; according to the method provided by the invention, resource occupation of model training is reduced, and the efficiency and precision of model prediction are improved.

Description

Technical Field

[0001] The present invention relates to a feature engineering method and system, and particularly to a feature engineering method and system applicable to feature recalibration for load forecasting, belonging to the technical field of electric load forecasting and analysis. Background Art

[0002] With the development of the global economy and the progress of technology, the power system is undergoing unprecedented changes. The wide application of new energy, the rise of smart grid technology, and the enhancement of user-side response capabilities have all posed new challenges to the planning and operation of the power system. Load forecasting, as a key link to ensure the stable, safe, and economic operation of the power system, is particularly important in this context. Load forecasting refers to the process of estimating the power demand within a certain future time period. It is not only the basis for power dispatching departments to formulate power generation plans but also one of the important bases for power market transactions. Accurate load forecasting helps to optimize resource allocation, reduce operating costs, improve service quality, and provide decision-making support for coping with emergencies.

[0003] Currently, the mainstream technical routes for load forecasting are mainly divided into classical statistical models and machine learning and artificial intelligence methods. Classical statistical models mainly perform time series analysis and regression analysis on time series data, including autoregressive methods, autoregressive moving average models, statistical-based analysis models, linear regression analysis, and nonlinear multiple regression analysis, etc.; machine learning and artificial intelligence-based methods mainly include artificial neural network models, recurrent neural network models, support vector machines, decision tree models, and ensemble learning models, etc. However, these methods are only limited to load forecasting in specific scenarios and cannot be effectively adjusted for load forecasting in different regions or under extreme weather conditions. Moreover, load forecasting has certain fuzzy logic and uncertainty problems, which pose higher requirements for the model. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a feature engineering method and system applicable to feature recalibration for load forecasting that can improve the accuracy of load forecasting.

[0005] Technical Solution: A feature engineering method applicable to feature recalibration for load forecasting according to the present invention includes:

[0006] Step 1: Obtain the historical load data and original meteorological data of the target area, and unify the time spans of the historical load data and the original meteorological data;

[0007] Step 2: Concatenate and perform anomaly detection on the historical load data and the original meteorological data after unifying the time span, and manually add Gaussian white noise perturbation after the detection to construct an initial sample set;

[0008] Step 3: Construct basic features based on the initial sample set, perform feature combination on the basic features, and use the wrapper and filter methods respectively to screen the features generated by the feature combination to generate a feature matrix;

[0009] Step 4: Input the feature matrix into the SE-Net model for feature matrix compression and remapping to obtain a feature sample set;

[0010] Step 5: Concatenate the initial sample set and the feature sample set to construct a training data set, and use the LSTM model for fitting, validation, and prediction based on the training data set.

[0011] Further, Step 2 includes: using 3-sigma anomaly detection for historical load data, using z-score anomaly detection for historical meteorological data, removing abnormal data and filling the data using the exponential smoothing method, and adding Gaussian white noise to the time series to construct an initial sample set.

[0012] Further, Step 3 of constructing basic features based on the initial sample set and performing feature combination on the basic features includes:

[0013] Obtain the load characteristics within several days before the day at each granularity time node of the initial sample set;

[0014] Weather characteristics obtained through combined operations based on the original meteorological data in the initial sample set; the combined operations include calculating the maximum value, minimum value, and average value, as well as using triangular transformation and feature operations;

[0015] Obtain the time series information and holiday information of the initial sample set, and use one-hot encoding to obtain time series features.

[0016] Further, Step 3 of using the wrapper and filter methods respectively to screen the features generated by the feature combination to generate a feature matrix includes:

[0017] Use the filter method to measure and evaluate the correlation degree between each generated feature and the prediction target using the Pearson pearson correlation coefficient and the maximum mutual information coefficient MIC;

[0018] Use the wrapper method to retain the optimal feature subset for the entire prediction task through the LVM feature selection strategy;

[0019] Fuse the features screened by the wrapper and filter methods to obtain a feature matrix.

[0020] Further, the use of the filter method measures and evaluates the correlation degree between each generated feature and the prediction target by using the Pearson correlation coefficient and the maximum information coefficient MIC. The calculation formula of the Pearson correlation coefficient is as follows:

[0021]

[0022] Among them, ρX,Y represents the covariance between two variables X and Y divided by the product of their respective standard deviations, which is used to reflect the linear correlation degree between variables. The greater the absolute value of ρX,Y, the stronger the correlation; the calculation formula of the maximum information coefficient MIC is as follows:

[0023]

[0024] Among them, f(x) and g(y) respectively represent the discretization processing of variables X and Y, I(f(x) and g(y)) represents the mutual information between the discretized X and Y, and H(f(x)) and H(g(y)) represent the entropy of the discretized X and Y.

[0025] Further, the step 4 includes:

[0026] Compress and reduce the dimension according to the spatial dimension of the feature matrix, perform feature encoding on the feature matrix of each dimension, and convert it into the global eigenvalue Z corresponding to the number of dimensions. The conversion formula is as follows: c , the conversion formula is as follows:

[0027]

[0028] Among them, H and W respectively represent the length and width of the feature matrix, i and j respectively represent the traversal values of H and W, u c and z c respectively represent the feature matrix sequences before and after the transformation;

[0029] Obtain the global eigenvalue and its relationship with different feature dimensions. First, perform dimensionality reduction and activation operations through the fully connected layer and the ReLU layer, and then normalize each global eigenvalue through the Sigmoid layer to obtain the activation values of different dimensions. Update the weight values corresponding to different feature matrices by introducing the time series weighted loss function. The formula is as follows:

[0030]

[0031] Among them, y n is the nth real load sample, is the corresponding predicted load, N represents the total sample size, ω n is the weight of the nth sample;

[0032] Obtain the activation value, and weight the activation value dimension by dimension onto the feature matrix before the initial input in the manner of matrix multiplication to complete the feature remapping of each feature matrix, and recalibrate the importance of the corresponding features of each feature matrix for the final prediction target to obtain the feature sample set.

[0033] Based on the same inventive concept, the present invention also provides a feature engineering system applicable to feature recalibration for load prediction, including:

[0034] A data acquisition module, configured to acquire historical load data and original meteorological data of a target area, and unify the time spans of the historical load data and the original meteorological data;

[0035] A data processing module, configured to splice and perform anomaly detection on the historical load data and the original meteorological data after unifying the time spans, and manually add Gaussian white noise perturbation after detection to construct an initial sample set;

[0036] A feature construction module, configured to construct basic features based on the initial sample set, perform feature combination on the basic features, and respectively perform feature screening on the features generated by the feature combination using wrapper and filter methods to generate a feature matrix;

[0037] A feature compression module, configured to input the feature matrix into the SE-Net model for feature matrix compression and remapping to obtain a feature sample set;

[0038] A model training and prediction module, configured to splice the initial sample set and the feature sample set to construct a training data set, and based on the training data set, use the LSTM model for fitting, verification and prediction.

[0039] Based on the same inventive concept, the present invention also provides a computer program product, including a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the feature engineering method applicable to feature recalibration for load prediction according to any one of the above are implemented.

[0040] Based on the same inventive concept, the present invention also provides a computing device, including: one or more processors, one or more memories, and one or more programs, the programs are stored in the memory and are configured to be executed by the processor, and when the programs are loaded into the processor, the steps of the feature engineering method applicable to feature recalibration for load prediction according to any one of the above are implemented.

[0041] Based on the same inventive concept, the present invention also provides a storage medium storing a computer program, the computer program including program instructions which, when executed by a processor, cause the processor to execute the steps of the feature engineering method for load prediction feature recalibration according to any one of the above.

[0042] Advantageous effects: Compared with the prior art, by optimizing device-to-device communication, adjusting the transmission power and the priority of users, the present invention balances the traffic pressure of different cells while enabling some incommunicable devices to still transmit data to the base station; the method proposed by the present invention achieves higher throughput and lower latency while guaranteeing the QoS of the primary service, and has higher applicability, stability and efficiency. Description of the Drawings

[0043] Figure 1 is the flowchart of the method according to an embodiment of the present invention;

[0044] Figure 2 is the schematic structural diagram of the SE-Net model according to an embodiment of the present invention;

[0045] Figure 3 is the schematic diagram of comparison of prediction results according to an embodiment of the present invention;

[0046] Figure 4 is the schematic diagram of comparison of load prediction accuracies between the recalibrated features and the features screened by other feature engineering according to an embodiment of the present invention. Detailed Embodiments

[0047] In order to enable those skilled in the art of the present technology to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application.

[0048] As shown in the Figure 1 accompanying drawings, the feature engineering method for load prediction feature recalibration in this embodiment includes:

[0049] Step 1: Obtain the historical load data and the original meteorological data of the target area, and unify the time spans of the historical load data and the original meteorological data;

[0050] Step 2: Concatenate and perform anomaly detection on the historical load data and the original meteorological data after unifying the time span, and manually add Gaussian white noise perturbation after the detection to construct an initial sample set;

[0051] Step 3: Construct basic features based on the initial sample set, perform feature combination on the basic features, and respectively perform feature screening on the features generated by the feature combination using the wrapper and filter methods to generate a feature matrix;

[0052] Step 4: Input the feature matrix into the SE-Net model for feature matrix compression and remapping to obtain a set of feature samples;

[0053] Step 5: Concatenate the initial sample set and the set of feature samples to construct a training data set, and based on the training data set, use the LSTM model for fitting, validation, and prediction.

[0054] Specifically, in Step 1, historical load data and original meteorological data from multiple cities under a certain province from January 1, 2021, to July 1, 2024, were obtained. The relevant original meteorological data includes temperature, humidity, wind speed, irradiance, air pressure, etc. The sample sampling granularity is one sample point every 15 minutes.

[0055] In Step 2, historical load data and original meteorological data with the same span were concatenated and anomaly detection was performed. After detection, Gaussian white noise perturbation was manually added to construct an initial sample set. Specifically:

[0056] Step 2.1: Delete some missing data in the historical load data and abnormal values such as Null or -999 in the original meteorological data.

[0057] Step 2.2: Perform 3-sigma detection on the historical load data. By calculating the mean and standard deviation of the historical load data, determine the data points outside 3 times the standard deviation from the mean in the normal distribution and list them as abnormal values. The 3-sigma detection formula is:

[0058]

[0059] where X is a random variable, μ is the expectation of the random variable, and the variance is σ 2 , when the variance of the variable X is smaller, the probability of P is greater;

[0060] For the original meteorological data, the z-score method is used for anomaly detection. The z-score method can manually correct the threshold setting. Since there are extreme weather conditions such as typhoons and cold snaps in the original meteorological data, and this type of data is easily misjudged as abnormal data by other algorithms, this method is used to increase the flexibility of anomaly data detection; its calculation formula is as follows:

[0061]

[0062] where T is the constant term setting, which can be dynamically corrected according to the dataset distribution. In this case, the T value is set to 0.675. Because the dataset is relatively sufficient, the threshold is set to 3, and mean and std represent the mean and standard deviation of the dataset respectively.

[0063] Step 2.3: Use the exponential smoothing method to fill in the historical load data and meteorological raw data after anomaly detection to ensure the integrity and continuity of the dataset in time series; at the same time, add Gaussian white noise to the time series to construct an initial sample set. White noise can simulate the errors in data acquisition and various possible disturbance factors in the real scenario. The filled data has problems such as simple linear changes or curve smoothing. Increasing the noise fluctuation can improve the anti-interference ability of the subsequent model when learning data and increase the model robustness.

[0064] In step 3, construct the basic features based on the initial sample set, perform feature combination on the basic features, and use the wrapper and filter methods to perform feature screening respectively to generate a feature matrix. Specifically:

[0065] Step 3.1: Perform feature engineering on the initial sample set after anomaly detection and filling, specifically including:

[0066] (1) Obtain the load values within several days before the day at each granularity time node as features. In this embodiment, the historical load data within one week is used as the load feature, specifically manifested as the load today t days ago at the same time point i, where t = 1,..., 6. i-t , t = 1,..., 6.

[0067] (2) Obtain the basic meteorological raw data such as temperature and wind speed, and add features such as the daily maximum temperature, minimum temperature, average temperature, and temperature before the day to the temperature data, and use triangular transformation and feature operations on this basis to obtain combined weather features;

[0068] In this embodiment, the method of feature combination is adopted. On the basis of the existing basic meteorological data such as temperature and wind speed provided, sum, difference, product, and quotient operations are performed between the basic weather features respectively to increase the number of combined features. Considering that the influence of temperature is relatively strong, features such as the daily maximum temperature, minimum temperature, and average temperature are added to the daily temperature data. And in the case of continuous high temperature or cooling, the load change will produce a cumulative effect, so the average temperature within the previous week is added as a feature supplement.

[0069] (3) Obtain the time series information and special festivals such as weekends and holidays, and use one-hot encoding for the time series features.

[0070] Step 3.2: Use the Pearson correlation coefficient and the maximum information coefficient MIC in the filter method to calculate the correlation degree between each column of relevant features obtained in step 3.1 and the target load to be predicted, and sort according to the correlation degree of different features.

[0071] The Pearson correlation coefficient adopted in this example has the following calculation formula:

[0072]

[0073] Among them, ρX,Y represents the covariance between two variables X and Y divided by the product of their respective standard deviations, which can reflect the degree of linear correlation between variables. The larger the absolute value of ρX,Y, the stronger the correlation.

[0074] In this example, the maximum information coefficient MIC is used, and its calculation formula is as follows:

[0075]

[0076] Among them, f(x) and g(y) respectively represent the discretization processing of variables X and Y, I(f(x); g(y)) represents the mutual information between the discretized X and Y, and H(f(x)), H(g(y)) represent the entropy of the discretized X and Y.

[0077] Step 3.3: Use the LVM (Las Vegas Wrapper) feature selection strategy in the wrapper method. This method randomly generates feature subsets from the feature set in Step 3.1 each time, adds the subsets to the model training, and uses cross-validation to evaluate the error of the feature subsets, and retains the feature subsets with smaller errors for the entire prediction task. The LVM method selects features through model feedback, so it can better capture the interaction between features. Different from the calculation method of the filter method independent of the algorithm, the wrapper method regards feature selection as a model optimization problem and evaluates the performance of different subsets during the model iteration process.

[0078] In this case, due to the large number of features, if all the generated features are used, it will not only increase the difficulty of model fitting, but also produce a large amount of feature redundancy. Therefore, the number of feature subsets is selected as k, and it is ensured that the dimension of the feature subset is consistent with the dimension of the features selected by the filter method. After splicing the features selected by the two methods, a feature space matrix is obtained.

[0079] In Step 4, the feature matrix is input into the SE-Net model for feature matrix compression and remapping to obtain a feature sample set, specifically:

[0080] Step 4.1: Compress the input k feature space matrices. Along the dimension of the feature matrix, each feature matrix is compressed into k 1x1 global feature values after being encoded by the global pooling layer (global polling). This feature value has the receptive field of the feature matrix before transformation and can reflect the representation ability of each feature matrix. Its calculation formula is as follows:

[0081]

[0082] where H and W represent the length and width of the feature matrix, respectively, and u c and z c represent the sequences of feature matrices before and after transformation, respectively.

[0083] Step 4.2: Further activate the obtained global eigenvalues to obtain the relationships between different eigenvalues. The SE-Net model structure is as Figure 2 shown, specifically including:

[0084] (1) Reduce the dimensionality of the k 1x1 global eigenvalues of the feature after passing through the fully connected layer;

[0085] (2) Then activate it through the ReLU activation function layer;

[0086] (3) Pass through the second fully connected layer to restore it to the original dimension;

[0087] (4) Pass through the sigmoid layer to normalize the k 1x1 global eigenvalues into k feature weight values. This weight value can be regarded as the importance representation of each feature matrix after feature compression. Introduce a time series weighted loss function to update the weight values corresponding to different feature matrices. The formula is as follows:

[0088]

[0089] where, y n is the nth true load sample, is the corresponding predicted load, N represents the total sample size, and ω n is the weight of the nth sample.

[0090] Step 4.3: Obtain k feature weight values. Through matrix multiplication, multiply them with the original feature space matrix one by one and weight them into the feature matrix. For the normalized feature weights, the unimportant part will tend to 0, and the result after matrix multiplication will also make the feature matrix tend to 0, completing the re-mapping and re-calibration of the original feature matrix, so that the model can focus on learning the features helpful for the prediction target and improve the discrimination ability of the model.

[0091] In Step 5, splice the initial sample set and the feature sample set, construct a training data set based on the training data set, and use the LSTM model for fitting, verification, and prediction. Specifically:

[0092] Determine that the input dimension of the model is (96, k), and the output dimension is (96, 1), where k represents the number of feature columns, which is obtained by screening in Step 3.

[0093] The present invention selects the LSTM model as the prediction model, and its training process is as follows:

[0094] (1) Take out the corresponding training data set, determine the input and output dimensions of the model, and use the ten-fold cross-validation method to repeat the following steps for multiple training times;

[0095] (2) Determine the weight initialization method as He initialization, and select a fixed learning rate of 10e-4;

[0096] (3) Select the mean squared error as the loss function, and the calculation formula is as follows:

[0097]

[0098] where m represents the number of sample points, y i represents the actual load value of the i-th sample, represents the model prediction of the i-th sample;

[0099] (4) The data set is propagated forward in the LSTM model for training calculation;

[0100] (5) The model is propagated backward to update the weight values;

[0101] Use the Adam gradient descent optimization algorithm to minimize the loss function MSE and obtain the trained model.

[0102] The trained LSTM model is used to predict the test data set, and the final prediction result is output. After obtaining the prediction result, the actual load data obtained subsequently is used to return to step 4.2 to update the feature weights and remap the features.

[0103] To verify the effectiveness of the method, the experiment will conduct a comparative analysis from two aspects: model comparison and feature comparison. Figure 3 This is the comparison diagram of the SE-Net+LSTM model of this embodiment with the LSTM model and the XGBoost model used alone. Figure 4 This is the schematic diagram of the comparison of the prediction accuracy of the recalibrated features of the method of this embodiment with the features screened by other feature engineering. It can be seen that the method of the present invention has obvious advantages both in terms of the model and the features. After feature recalibration, the model prediction accuracy can reach 98%.

[0104] Based on the same inventive concept, this embodiment also provides a feature engineering system applicable to load prediction feature recalibration, including:

[0105] A data acquisition module for acquiring the historical load data and meteorological raw data of the target area and unifying the time spans of the historical load data and the meteorological raw data;

[0106] A data processing module for splicing and anomaly detection of historical load data and raw meteorological data after a unified time span, and manually adding Gaussian white noise perturbations after detection to construct an initial sample set;

[0107] A feature construction module for constructing basic features based on the initial sample set, performing feature combination on the basic features, and respectively using wrapper and filter methods to screen the features generated by the feature combination to generate a feature matrix;

[0108] A feature compression module for inputting the feature matrix into the SE-Net model for feature matrix compression and remapping to obtain a feature sample set;

[0109] A model training and prediction module for splicing the initial sample set and the feature sample set to construct a training data set, and based on the training data set, using the LSTM model for fitting, verification and prediction.

[0110] Based on the same inventive concept, this embodiment also provides a computer program product, including computer programs / instructions, which when executed by a processor implement the steps of the feature engineering method for load prediction feature recalibration according to any one of the above.

[0111] Based on the same inventive concept, this embodiment also provides a computing device, including: one or more processors, one or more memories, and one or more programs, the programs are stored in the memory and configured to be executed by the processor, and when the programs are loaded into the processor, they implement the steps of the feature engineering method for load prediction feature recalibration according to any one of the above.

[0112] Based on the same inventive concept, in one embodiment of the present invention, there is also provided a storage medium, specifically a computer-readable storage medium. The computer-readable storage medium is a memory device in a computer device for storing programs and data. When the program is executed by a processor, the processor is caused to execute the steps of the feature engineering method applicable to load prediction feature recalibration according to any one of the above. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and, of course, the extended storage medium supported by the computer device. The computer-readable storage medium provides a storage space, and this storage space stores the operating system of the terminal. And, in this storage space, there are also stored one or more instructions suitable for being loaded and executed by the processor. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of a power distribution system reliability assessment method and system considering instantaneous power outage events in the above embodiments.

[0113] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0114] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for realizing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0115] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means that implement the function specified in one or more of the processes and / or blocks Figure 1 one or more of the processes and / or blocks Figure 1 specified in the block or blocks.

[0116] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the function specified in one or more of the processes and / or blocks Figure 1 one or more of the processes and / or blocks Figure 1 specified in the block or blocks.

[0117] The above are only embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval of the application.

Claims

1. A feature engineering method applicable to load forecasting feature recalibration, characterized in that, Including: Step 1: Obtain the historical load data and original meteorological data of the target area, and unify the time spans of the historical load data and the original meteorological data; Step 2: Concatenate and perform anomaly detection on the historical load data and the original meteorological data after unifying the time span, and manually add Gaussian white noise perturbation after detection to construct an initial sample set; Step 3: Construct basic features based on the initial sample set, perform feature combination on the basic features, and use the wrapper and filter methods respectively to perform feature screening on the features generated by the feature combination to generate a feature matrix; Step 4: Input the feature matrix into the SE-Net model for feature matrix compression and remapping to obtain a feature sample set; Step 5: Concatenate the initial sample set and the feature sample set to construct a training data set, and based on the training data set, use the LSTM model for fitting, validation and prediction.

2. The feature engineering method for load prediction feature recalibration according to claim 1, wherein The said Step 2 includes: Using 3-sigma anomaly detection for the historical load data, using z-score anomaly detection for the original meteorological data, removing the anomaly data and using the exponential smoothing method for data filling, and at the same time adding Gaussian white noise to the time series to construct an initial sample set.

3. The feature engineering method applicable to load prediction feature recalibration according to claim 1, wherein The construction of basic features based on the initial sample set in Step 3 and the feature combination on the basic features include: Obtain the load features within several days before the day at each granularity time node of the initial sample set; Weather features obtained by performing combination operations based on the original meteorological data in the initial sample set; The combination operations include calculating the maximum value, the minimum value and the average value, as well as using triangular transformation and feature operations; Obtain the time series information and holiday information of the initial sample set, and use one-hot encoding to obtain time series features.

4. The feature engineering method applicable to feature recalibration for load forecasting according to claim 1, wherein The use of the wrapper and filter methods respectively to perform feature screening on the features generated by the feature combination in Step 3 to generate a feature matrix includes: Using the filter method to measure and evaluate the correlation degree between each generated feature and the prediction target by using the Pearson correlation coefficient and the maximum mutual information coefficient MIC; Using the wrapper method, through the LVM feature selection strategy, retain the optimal feature subset for the entire prediction task; Fuse the features screened by the wrapper and filter methods to obtain a feature matrix.

5. The feature engineering method applicable to feature recalibration for load forecasting according to claim 4, characterized in that, The use of the filter method to measure and evaluate the correlation degree between each generated feature and the prediction target by using the Pearson correlation coefficient and the maximum mutual information coefficient MIC, where the calculation formula of the Pearson correlation coefficient is as follows: Among them, ρX,Y represents the covariance between two variables X and Y divided by the product of their respective standard deviations, which is used to reflect the linear correlation degree between variables. The greater the absolute value of ρX,Y, the stronger the correlation; The calculation formula of the maximum mutual information coefficient MIC is as follows: Among them, f(x) and g(y) respectively represent the discretization processes of variables X and Y. I(f(x) and g(y)) represents the mutual information between the discretized X and Y, and H(f(x)) and H(g(y)) represent the entropies of the discretized X and Y.

6. The feature engineering method applicable to load prediction feature recalibration according to claim 1, wherein The said step 4 includes: Compress and reduce the dimension according to the spatial dimension of the feature matrix, perform feature encoding on the feature matrix of each dimension, and convert it into the global eigenvalue Z corresponding to the number of dimensions. c , and the conversion formula is as follows: where H and W respectively represent the length and width of the feature matrix, i and j respectively represent the traversal values of H and W, and u c and z c respectively represent the sequences of feature matrices before and after transformation; Obtain the global eigenvalues and their relationships with different feature dimensions. First, perform dimensionality reduction and activation operations through a fully connected layer and a ReLU layer, and then normalize each global eigenvalue through a Sigmoid layer to obtain activation values of different dimensions. Update the weight values corresponding to different feature matrices by introducing a temporal weighted loss function. The formula is as follows: Among them, y n is the nth true load sample, is the corresponding predicted load, N represents the total sample size, ω n is the weight of the nth sample; Obtain the activation values, and weight each dimension of the activation values to the feature matrix before the initial input in the way of matrix multiplication to complete the feature remapping of each feature matrix, and recalibrate the importance of the corresponding features of each feature matrix to the final prediction target to obtain a feature sample set.

7. A feature engineering system applicable to load forecasting feature recalibration, characterized in that It includes: A data acquisition module, which is used to acquire the historical load data and meteorological raw data of the target area and unify the time spans of the historical load data and meteorological raw data; A data processing module, which is used to splice and perform anomaly detection on the historical load data and meteorological raw data after unifying the time span, and manually add Gaussian white noise perturbation after detection to construct an initial sample set; A feature construction module, which is used to construct basic features based on the initial sample set, perform feature combination on the basic features, and use wrapper and filter methods respectively to perform feature screening on the features generated by the feature combination to generate a feature matrix; A feature compression module, which is used to input the feature matrix into the SE-Net model to perform feature matrix compression and remapping to obtain a feature sample set; A model training and prediction module, which is used to splice the initial sample set and the feature sample set to construct a training data set, and based on the training data set, use the LSTM model for fitting, verification and prediction.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, the steps of the feature engineering method for feature recalibration applicable to load prediction according to any one of claims 1 to 6 are implemented.

9. A computing device, characterized in that, It includes: One or more processors, one or more memories, and one or more programs. The programs are stored in the memory and are configured to be executed by the processor. When the programs are loaded into the processor, the steps of the feature engineering method for feature recalibration applicable to load prediction according to any one of claims 1 to 6 are implemented.

10. A storage medium, characterized in that, The storage medium stores a computer program, and the computer program includes program instructions. When the program instructions are executed by the processor, the processor is enabled to execute the steps of the feature engineering method for feature recalibration applicable to load prediction according to any one of claims 1 to 6.

Images