KAN-BiLSTM power load prediction method based on composite factor construction

Abstract

The invention relates to the technical field of electric power engineering, and discloses a KAN-BiLSTM power load prediction method based on composite factor construction, and the method comprises the steps: carrying out the minimum and maximum normalization preprocessing of original data; a Gaussian mixture model is adopted to cluster daily load curve feature vectors, multiple load modes are divided, a composite factor is constructed through weighted linear combination based on an original influence factor and a Pearson's correlation coefficient absolute value of a power load, then an optimal feature set is screened out through a maximum information coefficient algorithm, and the optimal feature set is obtained. And independently training a KAN-BiLSTM model for each load mode, carrying out advanced nonlinear mapping by using a KAN layer based on B-spline parameterization, calling the corresponding model to obtain a predicted value, and then executing anti-normalization processing. According to the method, through combination of mode differentiation, composite feature enhancement and a high-interpretability model, the prediction precision in a complex scene is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric power engineering, in particular to a KAN-BiLSTM power load prediction method based on composite factor construction. BACKGROUND

[0002] Short-term load forecasting is the basis for real-time scheduling and power allocation of power systems, and its accuracy plays a key role in the economic operation and stability of the power grid. However, the current short-term load forecasting method still has significant shortcomings in capturing complex time series patterns, increasing prediction efficiency and accuracy.

[0003] The main challenge of the prior art is that a single model cannot achieve ideal prediction results, because power load data is affected by many nonlinear and non-stationary external factors such as weather and calendar types, resulting in a variable pattern and highly complex time series dependence of the load curve. In order to cope with this complexity, although combined prediction models have been widely used, they have advantages in improving prediction accuracy, but also bring significant increases in training time and usage costs, and the complexity of the model structure, which to some extent affects its deployment efficiency in actual systems.

[0004] Specifically, traditional neural networks such as multilayer perceptron (MLP) have a huge amount of parameters in their fully connected layers when processing high-dimensional input data, resulting in low efficiency of the model in processing high-dimensional data. In addition, traditional recurrent neural networks, especially when processing long sequences, are prone to problems such as gradient vanishing or gradient explosion, further limiting their effective capture of long-term dependencies. Therefore, existing methods have not been able to systematically optimize features and model architecture to balance prediction accuracy, parameter efficiency and computational cost, and a new method is needed to solve the problem of traditional prediction models in capturing complex time series patterns, improving accuracy while taking into account computational cost through efficient feature enhancement and lightweight, efficient neural network architecture. SUMMARY

[0005] In view of the shortcomings of the prior art, the present application provides a KAN-BiLSTM power load prediction method based on composite factor construction, which solves the problems of weak generalization ability of single models, low feature representation efficiency, and difficulty in balancing model accuracy and interpretability in existing power load prediction.

[0006] To achieve the above purpose, the present application is implemented by the following technical scheme: a KAN-BiLSTM power load prediction method based on composite factor construction.

[0007] The present application provides a KAN-BiLSTM power load prediction method based on composite factor construction in the first aspect, the steps of which include: Data preprocessing: data cleaning, outlier correction and missing value filling are performed on the original data set containing historical power load data and various original influence factor data, and then min-max normalization processing is performed, and the minimum value used for normalization is recorded and maximum value , to obtain a preprocessed data set.

[0008] Load pattern division: using the power load data in the preprocessed data set, the power load data arranged by natural days is transformed into day load curve feature vectors, a Gaussian mixture model (GMM) is used to cluster the set of day load curve feature vectors, at least two load patterns are divided, and the historical data is labeled with a pattern label, to obtain a data set with a pattern label.

[0009] The GMM clustering fits a predetermined number of multi-dimensional Gaussian distributions by an expectation maximization (EM) algorithm; for any day load curve feature vector , the posterior probability generated by the Gaussian distribution is calculated , and is attributed to the load pattern corresponding to the Gaussian distribution with the maximum posterior probability.

[0010] Pattern-specific optimal feature set generation: for each divided load pattern, a corresponding optimal feature set is independently generated using the data labeled with the pattern label, and the step of generating the optimal feature set includes: Correlation and composite factor construction: calculate the Pearson correlation coefficient between each original influence factor and the power load data in the pattern . According to the Pearson correlation coefficient, select a first original influence factor and a second original influence factor , and generate a composite factor

[0011] by weighted linear combination, wherein the weights of the weighted linear combination are determined by the absolute values of the Pearson correlation coefficients of the influence factors. wherein , , and are the Pearson correlation coefficients and between the first original influence factor and the second original influence factor

[0012] Feature screening: the maximum information coefficient (MIC) algorithm is used to screen features from the candidate feature set composed of the original influence factors and the composite factors, to form the optimal feature set. The screening strategy is to calculate the maximum information coefficient value between each feature and the power load sequence, and select all features greater than a preset threshold, or select the top

[0013] KAN-BiLSTM model training: for the divided load patterns, an independent KAN-BiLSTM model is trained for each corresponding optimal feature set, to obtain a group of trained models.

[0014] The KAN-BiLSTM model includes a BiLSTM network and a KAN layer. The BiLSTM network is used to receive the time series features extracted by the optimal feature set and output a combined hidden state vector ; the KAN layer is used to receive , and generates the normalized power load prediction value through a nonlinear mapping operation composed of learnable one-dimensional functions inside the KAN layer.

[0015] All learnable one-dimensional functions inside the KAN layer are parameterized by B-splines, and the nonlinear mapping operation is learned by updating the coefficients of the B-splines during model training.

[0016] Load prediction and denormalization: according to the features of the day to be predicted, the trained classifier is used to determine the load pattern to which the day to be predicted belongs, the model corresponding to the load pattern is called from the group of trained models, and the feature vector constructed for the day to be predicted is input, to obtain the normalized power load prediction value . Then, the normalized power load prediction value is denormalized to obtain the final actual power load prediction value .

[0017] ; wherein, is the actual power load prediction value obtained after denormalization; is the normalized prediction value output by the KAN-BiLSTM model; and are the maximum and minimum values in the original power load dataset, respectively.

[0018] The core innovation of the present application lies in the divide-and-conquer pattern recognition architecture, the correlation-based composite feature enhancement, and the combination of high interpretability and fitting ability of the model.

[0019] ​Split-mode prediction architecture (mode division): GMM is used instead of traditional K-Means for clustering, which can more accurately identify the potential patterns of power load curves (such as weekdays, weekends, holidays, etc.) from the perspective of probability distribution, effectively solving the problem of large differences in data distribution under different load patterns. Through accurate mode division of historical data, high-purity sub-datasets are provided for subsequent feature engineering and model training.

[0020] Composite factor construction (feature enhancement): The invention constructs a weighted linear composite factor based on the absolute value of the Pearson correlation coefficient between the original influence factor and the load. This construction method can efficiently fuse multiple influence factors that have an impact on the load but may have redundancy or linear relationship into a new feature, and at the same time, through the absolute value of the correlation coefficient as the weight, give higher weight to the factor that has greater impact on the load. This significantly enhances the feature's ability to represent the load and the quality of the model's input.

[0021] Introduction and role of KAN-BiLSTM model: BiLSTM network: used to capture the bidirectional dependence and sequence features of the optimal feature set in the time dimension, effectively handling the inherent time series characteristics of load prediction.

[0022] KAN layer: uses KAN instead of the traditional fully connected layer in deep learning models. KAN parameterizes each neural connection through B-spline functions, making the model's nonlinear mapping more interpretable and having stronger fitting ability, overcoming the problems of traditional fully connected layer black-box and ReLU activation function easily leading to neuron death, enabling the model to more efficiently and transparently learn the complex nonlinear relationship between features and loads, ultimately improving prediction accuracy.

[0023] The invention provides a KAN-BiLSTM power load prediction method based on composite factor construction. It has the following beneficial effects: 1、The invention uses Gaussian Mixture Model (GMM) to perform probability clustering on daily load curve feature vectors, achieving accurate division of the potential patterns of historical load data. By independently training KAN-BiLSTM prediction models for each sub-load pattern, a divide-and-conquer prediction architecture is formed, effectively isolating the heterogeneity of data distribution under different patterns, significantly improving the model's fitting ability and generalization performance for specific load patterns, and ultimately improving the power load prediction accuracy in complex scenarios.

[0024] 2. This invention generates composite factors by weighted linear combination based on the absolute value of the Pearson correlation coefficient between the original influencing factors and the power load. This method efficiently aggregates multi-source information by giving greater weight to factors with high correlation, thereby strengthening the feature's ability to represent the load. Subsequently, feature selection is performed by combining the maximum information coefficient (MIC), ensuring that the optimal feature set of the input model has the characteristics of high information content and minimal redundancy, thus optimizing the model's convergence speed and prediction performance from the feature input end.

[0025] 3. This invention integrates a B-spline-parameterized KAN layer into the output of a BiLSTM network. The KAN layer achieves advanced nonlinear mapping of features through a learnable one-dimensional function, avoiding the black box nature of traditional fully connected layers and enhancing the fitting accuracy for complex nonlinear relationships. This structure, combined with BiLSTM's deep extraction capability of temporal features, enables the model to maintain high prediction accuracy while possessing structured interpretability at the neuron connection level, helping engineers understand the mechanism by which features affect prediction results. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the power load data preprocessing module of the present invention; Figure 3 This is a schematic diagram of the load mode division module structure of the present invention; Figure 4 This is a schematic diagram of the composite factor construction and optimal feature set selection module of the present invention; Figure 5 This is a schematic diagram of the KAN-BiLSTM model structure of the present invention. Detailed Implementation

[0027] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] See attached document Figure 1 and attached Figure 2 The power load forecasting system provided in this embodiment of the invention may include a front-end processing module comprising a data acquisition module, a data cleaning module, and a data normalization module.

[0029] The data acquisition module is configured to obtain an original dataset. The original dataset includes historical power load data within a preset time range and a plurality of influence factor data corresponding to time points of the historical power load data. The influence factor data includes temperature data, humidity data, wind speed data, date type data, and holiday information data.

[0030] The data cleaning module is connected to the output end of the data acquisition module and is configured to process the original dataset to generate a cleaned dataset. The processing process aims to correct outliers in the data and fill in missing values to ensure the integrity and accuracy of the data input into the model.

[0031] Specifically, the data cleaning module first performs outlier identification and correction on numerical data in the dataset. Outlier data points are identified by a preset statistical threshold (e.g., three standard deviations or quartile range). After identifying the outlier data points, a predetermined algorithm (e.g., mean substitution or linear interpolation) is used to correct the data points.

[0032] Subsequently, the data cleaning module fills in missing values in the dataset. Missing data points are located by detecting null values or invalid values in the dataset, and a preset filling strategy (e.g., using the mean of the previous and subsequent time points or calculating by interpolation) is used to fill in the data points.

[0033] The data normalization module is connected to the output end of the data cleaning module and is configured to perform scale transformation on numerical features in the cleaned dataset to generate a normalized dataset. This operation aims to eliminate the dimensional differences between different features.

[0034] The data normalization module uses the min-max normalization method, whose calculation formula is: ; wherein, is the normalized data value; is the original data value to be processed; is the minimum value in the feature column to which the original data belongs; is the maximum value in the feature column to which the original data belongs. When performing the normalization operation, the data normalization module is also configured to record and store the minimum value and the maximum value corresponding to each feature column for subsequent prediction result denormalization steps.

[0035] Referring to the accompanying Figure 3 After preprocessing the original data, the embodiments of the present application further include a step of dividing the daily load curve into patterns, which is sequentially performed by the daily load curve characterization module and the Gaussian mixture model clustering module.

[0036] a day load curve featureization module, which is connected to the output of the data normalization module. The module is used to transform the continuous power load data of each natural day (e.g., 24 or 96 time-sequenced load data points contained in a natural day) into a fixed-dimension feature vector. The feature vector is directly composed of all the time-sequenced load data points within the natural day, thus completely characterizing the shape and fluctuation characteristics of the single-day load curve.

[0037] a Gaussian mixture model clustering module, which is connected to the output of the day load curve featureization module. The module is used to receive the feature vector set of all historical dates and perform clustering analysis on the set, dividing the day load curves with similar morphological characteristics into the same category, each category corresponding to a load pattern.

[0038] Specifically, the Gaussian mixture model clustering module uses the Gaussian mixture model (GMM) algorithm to divide the feature vector set into a preset number of clusters. The value is a pre-configured system parameter, the value of which is determined according to the analysis results of applying the Bayesian information criterion (BIC) or the profile coefficient and other evaluation indicators to the historical data set, to identify the number of significantly different load patterns existing in the data set. When performing clustering, the module fits a multi-dimensional Gaussian distribution through the expectation maximization (EM) algorithm. The algorithm iteratively updates the parameters of each Gaussian distribution, including the mean vector, the covariance matrix, and the mixing weight, until the model parameters converge, so that the log-likelihood function value of all data points reaches the maximum. After the model training is completed, for any input feature vector (i.e., the day load curve of any day), the Gaussian mixture model clustering module calculates the posterior probability generated by each of the Gaussian distributions, and attributes the feature vector to the cluster corresponding to the Gaussian distribution with the maximum posterior probability. Finally, the module outputs a data set with a pattern label, in which each load data of a historical date is assigned a unique pattern identification. The data set with the pattern label is used for subsequent independent feature selection and model training processes for different load patterns.

[0039] Referring to the accompanying Figure 4 After completing the load pattern division, the embodiments of the present application further include the step of enhancing and screening the influence factors in each pattern, which is sequentially performed by the correlation analysis module, the composite factor generation module, and the optimal feature set selection module.

[0040] A correlation analysis module, the input of which is connected with the output of the Gaussian mixture model clustering module. The module is used to calculate the Pearson correlation coefficient between each original influence factor and the power load data under each divided load mode independently.

[0041] Specifically, for any original influence factor and power load , the module calculates the correlation coefficient of them. ; wherein, the Pearson correlation coefficient, the total number of data points (sample size), the i-th observation value of any original influence factor D, the i-th observation value of the power load , the average value (Mean) of all observation values of any original influence factor D, the average value (Mean) of all observation values of the power load .

[0042] A composite factor generation module, the input of which is connected with the output of the correlation analysis module. The module is used to select at least two original influence factors according to the calculated correlation coefficient, and generate a new composite factor through weighted linear combination.

[0043] The module first selects original influence factors and according to the preset rules, and obtains their corresponding Pearson correlation coefficients and . Then, the weights and are calculated, and the calculation formula is: ; ; wherein, and are the absolute values of the correlation coefficients and , respectively.

[0044] Finally, the module generates the composite factor through the following formula: ; wherein, is the composite factor value, ,​​​ for , The value of the corresponding factor, , They are respectively , The weight, and , The sum is 1.

[0045] This module will generate one or more composite factors. As a new feature column, it is merged with the original set of influencing factors to form a candidate feature set. The optimal feature set selection module is connected to the output of the composite factor generation module. This module is used to select the subset of features with the strongest correlation to power load from the candidate feature set, which serves as the final input to the prediction model.

[0046] This module employs the Maximum Information Coefficient (MIC) algorithm. For each feature in the candidate feature set (including the original influencing factors and newly generated composite factors), this module calculates its maximum information coefficient value with respect to the electricity load sequence. The maximum information coefficient measures the strength of the linear and nonlinear relationship between two variables.

[0047] After the calculation is completed, the module determines the optimal feature set according to the preset screening strategy. The selection strategies include: selecting all features with a maximum information coefficient value greater than a preset threshold, or selecting the top features with the maximum information coefficient value. The characteristics, among which This is a pre-configured integer parameter. The module ultimately outputs the optimal feature set. This will be fed into subsequent prediction models.

[0048] See attached document Figure 5 The BiLSTM network serves as a feature extraction module, with its input connected to the output of the optimal feature set selection module. It is used to perform deep temporal feature extraction on the input time series feature data.

[0049] The BiLSTM network consists of a forward Long Short-Term Memory (LSTM) network layer and a backward Long Short-Term Memory (LSTM) network layer running in parallel. Both network layers receive the same input, namely the optimal feature set. Optimal feature set Constructed into a system with a preset time step The sequence data is in the form of ,in Is An input vector that contains multiple features at any given time.

[0050] Forward LSTM network layers according to arrive processing the input sequence in time order . At each time step , the computing unit of the layer receives the input vector at the current time step and the hidden state at the previous time step , updates its cell state and computes the forward hidden state at the current time step through its internal input gate, forget gate and output gate mechanism. This gating mechanism enables the network to selectively memorize and forget the history information, thus capturing the long-term dependencies in the time series.

[0051] processing the input sequence in time reverse order . At each time step , the computing unit of the layer receives the input vector at the current time step and the hidden state at the next time step , and computes the backward hidden state at the current time step with the same internal mechanism as the forward LSTM network layer. This reverse processing procedure is used to capture the influence of the future context information on the current time step. At each time step , the BiLSTM network concatenates the forward hidden state output by the forward LSTM network layer and the backward hidden state output by the backward LSTM network layer to generate the combined hidden state at the time step. The computation process is as follows: ; where denotes the vector concatenation operation. If the number of hidden units of the forward and backward LSTM is both , then the dimension of the combined hidden state is .

[0052] The BiLSTM network finally outputs a hidden state sequence . This sequence fuses the past and future context information of the input sequence at each time point. In the embodiments of the present application, the combined hidden state at the last time step is taken as the output of the entire BiLSTM network and is input into the subsequent KAN layer for processing. The number of layers of the BiLSTM network and the number of hidden units of each LSTM unit are both preset model hyperparameters.

[0053] The KAN layer in the KAN-BiLSTM model is used as a prediction output module, and an input end of the KAN layer is connected with an output end of the BiLSTM network. The KAN layer is used to replace a full-connection output layer in a traditional neural network, and is used for nonlinear mapping of deep time sequence features extracted by the BiLSTM network, so as to generate a final power load prediction value.

[0054] The input of the KAN layer is a combined hidden state vector output by the BiLSTM network at a last time step. The core of the KAN layer is that a connection between neurons of the KAN layer is not a combination of a traditional weight and a fixed activation function, but is defined by a one-dimensional function that can be learned in a network training process. Specifically, each one-dimensional function that can be learned in the KAN layer is parameterized by a spline function.

[0055] In an embodiment of the present application, the function is constructed by using B-splines. The construction process includes: first, a fixed grid composed of multiple nodes is divided on an input domain of the function; second, a group of B-spline basis functions is defined on the grid; and finally, a specific shape of the spline function is determined by linear combination of the B-spline basis functions through a group of learnable coefficients (i.e., control points).

[0056] In a forward propagation process of the model, the combined hidden state vector is fed into the KAN layer. The KAN layer performs a composite function operation on the input vector through a plurality of learnable functions parameterized by B-splines in the KAN layer, and maps the high-dimensional feature vector into a single scalar output value . The process can be represented as: ; wherein represents a nonlinear mapping operation composed of the learnable spline function, and in a back propagation training process of the model, the B-spline coefficients used for defining the shape of the spline function will be iteratively updated according to the gradient of the loss function. In this way, the KAN layer can automatically learn and form an optimal nonlinear mapping relationship to fit the complex correlation between the input features and the load value. The scalar value output by the KAN layer is a normalized power load prediction value. The prediction value is then transmitted to a prediction result denormalization module for processing.

[0057] After the structure of the KAN-BiLSTM model is defined, an embodiment of the present application further includes a step of training and optimizing the model, which is cooperatively executed by a loss function calculation module and a parameter optimization module.

[0058] In this embodiment, a complete model training process is executed independently for each previously defined load pattern. Specifically, the dataset with pattern labels corresponding to that load pattern is first divided into a training set and a validation set according to a preset ratio. During training, the model is trained iteratively. In each iteration, a batch of sample data is first selected from the training set. This batch of data contains multiple time-series samples, each of which is composed of the optimal feature set. and its corresponding actual power load value constitute.

[0059] Batch data is input into a defined KAN-BiLSTM model for a forward propagation calculation. The model outputs a normalized prediction value for each sample in the batch. The loss function calculation module receives the batch predicted values ​​output by the model and the actual loading values ​​in the batch samples, and calculates the error between the two. In this embodiment, the mean squared error (MSE) is used as the loss function. The calculation formula is as follows: ; in, For the first The true loading value of each sample; For the model to the first The predicted value output for each sample; This represents the number of samples in this batch.

[0060] The parameter optimization module is used to calculate the error value output by the module based on the loss function. This involves adjusting all trainable parameters in the KAN-BiLSTM model. In this embodiment, the Adam optimizer is used to perform this function.

[0061] This module first calculates the loss function using the backpropagation algorithm. The gradient is calculated relative to the gradient of each trainable parameter in the model (including the weight parameters of the BiLSTM network layer and the B-spline coefficients of the KAN layer). Subsequently, the Adam optimizer updates the values ​​of all trainable parameters based on the calculated gradients, combined with its internally maintained first-order and second-order moment estimates, to optimize the loss function. The value is adjusted to decrease.

[0062] The complete process of selecting batches, forward propagation, calculating loss, back propagation, and updating parameters will be repeated until all samples in the training set are traversed, which completes an epoch. The entire training process will be repeated for a predetermined number of training epochs until the model's performance converges on the validation set.

[0063] Learning rate in model training process Batch size and the total number of training epochs are all pre-configurable hyperparameters. Ultimately, for each load pattern, an independent and trained KAN-BiLSTM model is generated and stored for subsequent load prediction stage calls.

[0064] After completing the KAN-BiLSTM model training for different load patterns, the embodiments of the present application further include a process of applying the trained model for actual load prediction, which is sequentially executed by a mode determination module, a model loading module, and a prediction execution module.

[0065] The mode determination module is used to receive various influence factor data (e.g., weather forecast data and date type data) of the day to be predicted and determine the load pattern to which the day to be predicted belongs according to the data. In this embodiment, the module contains a trained classifier internally. The classifier is trained in the model training stage with the influence factors in the historical data as input and the pattern labels output by the Gaussian Mixture Model clustering module as output, so as to learn the mapping relationship between the influence factors and the load patterns. In the prediction stage, the influence factor data of the day to be predicted is input to the classifier, and the output of the classifier is the load pattern identification to which the day to be predicted belongs.

[0066] The model loading module is connected to the output end of the mode determination module. The module retrieves and loads the previously independently trained KAN-BiLSTM model corresponding to the received load pattern identification from the model storage.

[0067] The prediction execution module is connected to the output end of the model loading module. The module first performs the same data processing operations as in the model training stage on the received influence factor data of the day to be predicted. Specifically, it includes:

[0068] ​​​​​

[0069] According to the optimal feature set determined under the load mode, a feature vector in time series format meeting the input requirements of the model is constructed from the normalized impact factor and the generated composite factor.

[0070] Subsequently, the module feeds the constructed feature vector as input into the loaded KAN-BiLSTM model. The model performs a complete forward propagation calculation, and finally outputs a scalar value, which is the normalized power load prediction value . The prediction value is then transmitted to the prediction result denormalization module.

[0071] The embodiment of the present application further includes a step of denormalizing the normalized prediction value output by the prediction execution module, which is performed by the prediction result denormalization module.

[0072] The prediction result denormalization module is connected to the output end of the prediction execution module. The module is used to transform the received power load prediction value in normalized scale (for example, in the interval of 0 to 1) back to its original physical unit and value range.

[0073] Specifically, the module first retrieves the minimum value and the maximum value corresponding to the power load data column recorded and stored when the data normalization module performs initial data processing from the data storage unit. Subsequently, the prediction result denormalization module performs an inverse scale transformation operation on the received normalized prediction value using the retrieved values. The calculation formula of the operation is: ; wherein, is the actual power load prediction value obtained after denormalization processing; is the normalized prediction value output by the KAN-BiLSTM model; and are the maximum value and the minimum value in the original power load data set, respectively.

[0074] The actual power load prediction value output by the module is the final result of the power load prediction method provided by the embodiment of the present application.

Claims

1. A KAN-BiLSTM power load forecasting method based on composite factor construction, characterized in that, Includes the following steps: The raw dataset containing historical power load data and various raw influencing factor data is preprocessed to obtain a preprocessed dataset. Using the power load data in the preprocessed dataset, the power load data arranged by natural day is transformed into daily load curve feature vectors, and a Gaussian mixture model is used to cluster the set of daily load curve feature vectors to divide them into multiple load patterns, thereby labeling the historical data in the preprocessed dataset with pattern labels. For each identified load pattern, an optimal feature set is independently generated using data labeled with the pattern. The steps for generating the optimal feature set include: Calculate the Pearson correlation coefficient between each original influencing factor and the power load data under the described model; Based on the Pearson correlation coefficient, multiple original influencing factors are selected and a composite factor is generated by a weighted linear combination, wherein the weight of the weighted linear combination is determined by the absolute value of the Pearson correlation coefficient of the at least two original influencing factors. The maximum information coefficient algorithm is used to select features from the original influencing factors and the composite factors to form the optimal feature set; For each load pattern, a KAN-BiLSTM model is independently trained using the corresponding optimal feature set, resulting in a set of trained KAN-BiLSTM models. Based on the characteristics of the day to be predicted, the load pattern to which the day to be predicted belongs is determined. The model corresponding to the load pattern in the set of trained KAN-BiLSTM models is called, and the feature vector constructed for the day to be predicted is input. After obtaining the normalized power load prediction value, the normalized power load prediction value is subjected to inverse normalization processing to obtain the final actual power load prediction value.

2. The KAN-BiLSTM power load forecasting method based on composite factor construction according to claim 1, characterized in that, The step of clustering the daily load curve feature vector set using a Gaussian mixture model includes: Fit the preset quantity using the expectation-maximization algorithm. A multidimensional Gaussian distribution is used until the model parameters converge; For any input daily load curve feature vector, calculate the value derived from the above. Each posterior probability is generated by a Gaussian distribution, and the feature vector is assigned to the load pattern corresponding to the Gaussian distribution with the largest posterior probability, thus obtaining a dataset with pattern labels.

3. The KAN-BiLSTM power load forecasting method based on composite factor construction according to claim 1, characterized in that, The step of generating composite factors through weighted linear combination includes: Select the first and second original impact factors and obtain the corresponding Pearson correlation coefficients; Calculate the weight of the first original impact factor, whereby the weight is the ratio of the absolute value of the Pearson correlation coefficient corresponding to the first original impact factor to the sum of the absolute values ​​of the Pearson correlation coefficients corresponding to the two original impact factors. Calculate the weight of the second original impact factor, whereby the weight is the ratio of the absolute value of the Pearson correlation coefficient corresponding to the second original impact factor to the sum of the absolute values ​​of the Pearson correlation coefficients corresponding to the two original impact factors. The first original impact factor is multiplied by the first weight, the second original impact factor is multiplied by the second weight, and the two products are added together to generate the composite factor.

4. The KAN-BiLSTM power load forecasting method based on composite factor construction according to claim 1, characterized in that, The step of using the maximum information coefficient algorithm to screen features from the original influence factors and the composite factors includes: Calculate the maximum information coefficient value between each feature in the candidate feature set and the power load sequence; Based on the maximum information coefficient value, select all features whose maximum information coefficient value is greater than a preset threshold, or select the top features with the highest maximum information coefficient value. The features of the bits constitute the optimal feature set, wherein It is a pre-configured integer.

5. The KAN-BiLSTM power load forecasting method based on composite factor construction according to claim 1, characterized in that, The KAN-BiLSTM model includes a BiLSTM network and a KAN layer; The BiLSTM network is used to receive the optimal feature set, extract time series features through a forward long short-term memory network layer and a backward long short-term memory network layer, and output a combined hidden state vector. The KAN layer is used to receive the combined hidden state vector and generate the normalized power load forecast value through an internal nonlinear mapping operation consisting of a learnable one-dimensional function.

6. The KAN-BiLSTM power load forecasting method based on composite factor construction according to claim 5, characterized in that, All learnable one-dimensional functions within the KAN layer are parameterized using B-splines, and the nonlinear mapping operation is learned by updating the coefficients of the B-splines during model training.

7. The KAN-BiLSTM power load forecasting method based on composite factor construction according to claim 1, characterized in that, The steps for independently training a KAN-BiLSTM model include: Mean squared error is used as the loss function; The gradient of the loss function with respect to the trainable parameters of the model is calculated using the backpropagation algorithm, and the trainable parameters are updated using the Adam optimizer.

8. The KAN-BiLSTM power load forecasting method based on composite factor construction according to claim 1, characterized in that, The step of determining the load pattern based on the characteristics of the day to be predicted includes: The data of various influencing factors for the day to be predicted are input into a trained classifier, and the output of the classifier is the load pattern identifier of the day to be predicted.

9. The KAN-BiLSTM power load forecasting method based on composite factor construction according to claim 1, characterized in that, The preprocessing steps for the original dataset include: Clean the original dataset, correct outliers, and fill in missing values; Perform min-max normalization on the cleaned data and record the minimum and maximum values ​​used for normalization.

10. The KAN-BiLSTM power load forecasting method based on composite factor construction according to claim 1, characterized in that, The steps of performing denormalization on the normalized power load forecast include: Using the recorded minimum and maximum values, perform an inverse scaling transformation on the normalized power load forecast to obtain the final actual power load forecast.

Images