Multi-factor efficient fusion microgrid load short-term prediction model optimization method

Abstract

The invention relates to the technical field of micro-grid load prediction, and discloses a multi-factor efficient fusion micro-grid load short-term prediction model optimization method. According to the method, micro-grid load historical data and multi-source factor data including information such as meteorology and electricity price are collected, and one-hot coding and normalization processing are carried out. Load data is decomposed into high-frequency and low-frequency components by using a variational mode decomposition technology, and prediction is carried out by respectively adopting a long-short-term memory network integrated with an attention mechanism and a linear regression model. And by calculating the difference between the source domain data and the target domain data in the model hidden layer feature space, the domain adaptive capacity of the model is enhanced, and accurate short-term prediction of the microgrid load is realized.

Description

Technical Field

[0001] This invention relates to the field of microgrid load forecasting technology, specifically to an optimization method for a multi-factor efficient fusion microgrid load short-term forecasting model. Background Technology

[0002] In microgrid systems, short-term load forecasting is crucial for ensuring stable system operation and optimizing resource allocation. However, traditional forecasting methods often rely solely on historical load data, neglecting the influence of multiple factors such as weather, electricity prices, and distributed power generation output, resulting in limited forecast accuracy. Furthermore, the significant differences in data distribution between different microgrids weaken the model's transferability, making it difficult to directly apply to new scenarios.

[0003] Therefore, developing a short-term microgrid load forecasting method that can efficiently integrate multiple factors and has good domain adaptability is key to improving forecast accuracy. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the purpose of this invention is to provide an optimization method for a microgrid load short-term prediction model that efficiently integrates multiple factors, thus solving the problems of existing prediction methods failing to efficiently integrate multiple source factors and exhibiting poor adaptability.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides an optimization method for a microgrid load short-term prediction model that integrates multiple factors efficiently, the method comprising: Historical load data and multi-source factor data of microgrids are collected. The multi-source factor data is divided into categorical factors and numerical factors. One-hot encoding is performed on categorical factors and normalization is performed on numerical factors. The processed multi-source factor data is then aligned with the historical load data sequence according to the timestamp to construct the source domain dataset. Variational mode decomposition technology is used to decompose the aligned load history data sequence into K intrinsic mode function components; based on the center frequency value of each component, the first M components with higher center frequency values ​​are classified as high frequency components, and the remaining KM components are classified as low frequency components. Construct a combined prediction model, which includes: The high-frequency prediction sub-model employs a long short-term memory network with an integrated attention mechanism to predict each high-frequency component sequence. The attention mechanism performs feature weighted fusion of the multivariate factors input to the sub-model. The low-frequency prediction sub-model uses a linear regression model to predict each low-frequency component sequence. Based on the source domain dataset, the component sequences obtained from decomposition and the corresponding multi-source factor data are used as inputs. The combined prediction model is pre-trained with the goal of minimizing the mean square error between the predicted value and the true value of the component. The pre-trained combined prediction model is transferred to the target micronet, and the model parameters are fine-tuned using sample data from the target micronet. During the fine-tuning process, the maximum mean difference between the source domain and target domain data in the hidden layer feature space of the model is calculated and used as a regularization term to form the loss function together with the prediction error. For the period to be predicted, multi-source factor data is acquired in real time and processed, then input into the fine-tuned combined prediction model to obtain the predicted values ​​of high-frequency and low-frequency components respectively; the predicted values ​​of all components are summed and reconstructed to obtain the final short-term predicted value of microgrid load.

[0006] Preferably, in one possible implementation of the first aspect, the process of decomposing using variational mode decomposition technology includes: The number of intrinsic mode function components K is initially determined by calculating the power spectrum of the load history data sequence and identifying the number of significant spectral peaks in the power spectrum. Variational mode decomposition (VMD) is employed to adaptively decompose the historical load data sequence into K narrowband eigenmode function components by solving a constrained variational problem and introducing a quadratic penalty factor and Lagrange multipliers for iterative optimization.

[0007] Preferably, in one possible implementation of the first aspect, the process for identifying the significant spectral peaks is as follows: After smoothing the power spectrum, local maxima are searched in the frequency domain, and a dynamic threshold is set. Local maxima with power values ​​higher than the dynamic threshold are identified as significant spectral peaks.

[0008] Preferably, in one possible implementation of the first aspect, the high-frequency prediction sub-model employs a long short-term memory network and integrates a multi-head attention mechanism layer at its input end for feature weighted fusion of the input multi-source factors; The low-frequency prediction sub-model adopts a multiple linear regression model with L2 regularization; The high-frequency prediction sub-model and the low-frequency prediction sub-model are set up in parallel and are used to process the high-frequency intrinsic mode function components and the low-frequency intrinsic mode function components, respectively.

[0009] Preferably, in one possible implementation of the first aspect, the multi-head attention mechanism layer specifically includes: The multi-source factor data at each time step is mapped into query vector, key vector, and value vector through linear transformation; The vector is distributed to multiple attention heads for parallel computation. Each attention head calculates the dot product of the query vector and the key vector and scales it. Attention weights are obtained through the Softmax function, and then the value vectors are summed in a weighted manner. Finally, the outputs of all attention heads are concatenated and a linear transformation is applied to obtain the feature-weighted fusion result.

[0010] Preferably, in one possible implementation of the first aspect, the pre-training process includes: Each intrinsic mode function component sequence obtained by decomposition in the source domain dataset is combined with the corresponding multi-source factor data to form a training sample. With the goal of minimizing the total mean square error between the component prediction values ​​and the actual component values ​​output by the combined prediction model, the parameters of the high-frequency prediction sub-model and the low-frequency prediction sub-model are jointly optimized using the backpropagation algorithm.

[0011] Preferably, in one possible implementation of the first aspect, the fine-tuning process includes: The parameters of the pre-trained combined prediction model are adjusted using sample data from the target microgrid. During fine-tuning, the loss function consists of prediction error and a regularization term. The regularization term is calculated based on the maximum mean difference between the source and target domain data in the hidden feature space of the model. This difference is measured by a Gaussian kernel function to measure the distance between the feature distributions.

[0012] Preferably, in one possible implementation of the first aspect, in the calculation of the maximum mean difference, the bandwidth parameter of the Gaussian kernel function is adaptively selected according to the dimension of the feature space.

[0013] Preferably, in one possible implementation of the first aspect, the process of obtaining the short-term forecast value of the microgrid load includes: For the period to be predicted, multi-source factor data is collected in real time, and after one-hot encoding or normalization, it is input into the fine-tuned combined prediction model. The model outputs prediction sequences for high-frequency components and low-frequency components respectively. The predicted values ​​for high-frequency components are generated by the high-frequency prediction sub-model, and the predicted values ​​for low-frequency components are generated by the low-frequency prediction sub-model. The predicted values ​​of all components are summed point by point over time to reconstruct a complete load forecast sequence, which serves as the short-term load forecast value for the microgrid.

[0014] The beneficial effects of this invention are as follows: This invention collects and efficiently integrates historical load data and multi-source factor data of microgrids, uses variational mode decomposition technology to decompose the load data into high-frequency and low-frequency components, and uses a long short-term memory network with integrated attention mechanism and a linear regression model for prediction respectively, which effectively improves the prediction accuracy.

[0015] Meanwhile, by calculating the maximum mean difference between the source and target domain data in the hidden layer feature space of the model and incorporating it as a regularization term into the loss function, the domain adaptability of the model is enhanced, enabling it to quickly adapt to different microgrid environments and improving the versatility and practicality of the prediction method. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This application provides a flowchart of a method for optimizing a microgrid load short-term prediction model through efficient fusion of multiple factors. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: As Figure 1 As shown, this invention provides a method for optimizing a microgrid load short-term prediction model using a multi-factor efficient fusion approach. The method includes: Historical load data and multi-source factor data of microgrids were collected. The multi-source factor data was divided into categorical factors and numerical factors. One-hot encoding was performed on categorical factors and normalization was performed on numerical factors. The processed multi-source factor data was then aligned with the historical load data sequence according to the timestamp to construct the source domain dataset.

[0020] In this embodiment, the historical load data of the microgrid is sourced from the historical operation database of the target microgrid system. The collection period covers hourly records for multiple consecutive full years, ensuring sufficient temporal coverage and seasonal characteristics. Multi-source factor data is acquired synchronously from multiple independent systems, including a meteorological monitoring platform, a power grid dispatch center, and the microgrid's internal monitoring system. These factors are divided into two main types: categorical factors and numerical factors. Categorical factors include discrete states such as weather types (sunny, rainy, snow) and time attributes such as weekdays and holidays. Numerical factors involve continuous environmental parameters such as temperature, humidity, wind speed, and light intensity, as well as operational variables such as electricity price, distributed power output, and energy storage state of charge.

[0021] In the data preprocessing stage, one-hot encoding is used to transform categorical factors. Specifically, each categorical variable is mapped to a binary vector, where the length of the vector equals the number of possible categories for that factor. For example, the weather type factor includes three categories: sunny, rainy, and snowy. This is encoded into a three-bit binary vector, with sunny represented as 100, rainy as 010, and snowy as 001. This process eliminates the disordered differences between categories and transforms discrete data into numerical form. For numerical factors, the min-max normalization method is used to linearly scale the original data of each factor to the range of 0 to 1, eliminating the influence of different dimensions between factors, avoiding large-scale numerical factors dominating model training, and improving model convergence speed and prediction stability.

[0022] The processed multi-source factor data is timestamped and aligned with the historical load data series. Due to the diverse data sources and inconsistent collection frequencies, the time resolution granularity is first standardized to the hourly level. During alignment, the multi-source factor data is interpolated or aggregated using the timestamps of the historical load data as a benchmark, ensuring that each time point has a complete factor vector. For example, meteorological data, collected at the minute level, is aggregated into an hourly mean using an arithmetic mean; while electricity price data, updated daily, is expanded to the hourly granularity through forward padding. The aligned dataset constitutes a multi-dimensional time-series matrix, with each row corresponding to a time point, containing the load value and all processed factor data.

[0023] The final source domain dataset structure includes timestamp sequences, historical load sequences, and multi-source factor feature matrices. The dataset is divided into training and validation sets, with the training set used for model parameter learning and the validation set used for hyperparameter tuning to ensure the representativeness of the data distribution and the model's generalization ability.

[0024] Variational mode decomposition (VMD) is used to decompose the aligned load history data sequence into K intrinsic mode function (IMF) components. Based on the center frequency values ​​of each component, the first M components with higher center frequency values ​​are classified as high-frequency components, and the remaining KM components are classified as low-frequency components.

[0025] In this embodiment, variational mode decomposition (VMD) is used to decompose the aligned load history data sequence. First, the number K of intrinsic mode function components is initially determined by calculating the power spectral density of the load history data sequence. Specifically, a Fourier transform is performed on the load sequence to obtain the power spectrum, which is then smoothed to suppress noise interference. The smoothing method uses a moving average filter, with the window length adaptively selected based on the total sequence length. Local maxima of the smoothed power spectrum are searched in the frequency domain, and a dynamic threshold is set to identify significant spectral peaks. This dynamic threshold is calculated based on the overall statistical characteristics of the power spectrum; in this embodiment, it is the weighted sum of the mean and standard deviation of the power spectrum values. Local maxima with power values ​​higher than the dynamic threshold are identified as significant spectral peaks, and their number is the initial value of K. If the number of significant spectral peaks is too small or too large, fine-tuning is performed based on the temporal characteristics of the load sequence to finally determine the value of K.

[0026] After determining the value of K, variational mode decomposition (VMD) is used to adaptively decompose the historical load data sequence into K narrowband intrinsic mode function (EMF) components. VMD is achieved by solving a constrained variational problem, the goal of which is to decompose the original sequence into multiple mode functions such that the sum of the estimated bandwidths of each mode is minimized. Specifically, a quadratic penalty factor is introduced to enhance the robustness of the decomposition, and Lagrange multipliers are used to transform the constrained problem into an unconstrained optimization problem. Iterative optimization employs the alternating direction multiplier method, updating the mode functions and their center frequencies at each step until the convergence condition is met. The convergence criterion is set based on the relative rate of change of the mode functions; typically, iteration stops when the change in the Euclidean distance of the mode functions in consecutive iterations is less than a preset tolerance. Through this process, the load sequence is decomposed into K EMF components with different center frequencies, each component having a clear physical meaning, corresponding to different frequency components in the load variation.

[0027] After decomposition, the components are divided based on their center frequency values. The center frequency of each intrinsic mode function component is calculated, which is the centroid of the component energy in the frequency domain. The center frequency values ​​are arranged in descending order, and the first M components are classified as high-frequency components, while the remaining KM components are classified as low-frequency components. The value of M is determined based on the cumulative contribution rate of the frequency distribution. In this embodiment, a critical point is selected where the cumulative energy of the high-frequency components reaches 80% of the total energy. High-frequency components mainly capture short-term fluctuations and random noise in the load sequence, such as rapid changes caused by sudden weather changes or equipment start-ups and shutdowns; low-frequency components characterize the long-term trend and periodicity of the load, such as daily cycles or seasonal effects.

[0028] A combined prediction model is constructed, which includes: a high-frequency prediction sub-model, which uses a long short-term memory network with an integrated attention mechanism to predict each high-frequency component sequence, wherein the attention mechanism performs feature weighted fusion of the multivariate factors input to the sub-model; and a low-frequency prediction sub-model, which uses a linear regression model to predict each low-frequency component sequence.

[0029] In this embodiment, the constructed combined prediction model consists of two parallel sub-models: a high-frequency prediction sub-model and a low-frequency prediction sub-model. The high-frequency prediction sub-model processes high-frequency intrinsic mode function (IMF) components. It employs a long short-term memory (LSTM) network as its basic architecture and integrates a multi-head attention mechanism layer at the network input to achieve feature-weighted fusion of the input multi-source factor data. The low-frequency prediction sub-model targets low-frequency IMF components and uses a multiple linear regression model with L2 regularization. The two sub-models are structurally independent and functionally complementary, working together to complete the prediction task of the decomposed component sequences.

[0030] The long short-term memory (LSTM) network portion of the high-frequency prediction sub-model employs a multi-layered stacked structure, with each layer containing several memory units. Each memory unit contains an input gate, a forget gate, and an output gate to control the flow and storage of information. The network input consists of high-frequency component sequences and their corresponding multi-source factor data, with the sequence length determined based on the historical time window. The LTM network captures long-term dependencies in the sequence through a gating mechanism, effectively modeling the nonlinear dynamic characteristics of high-frequency components. A multi-head attention mechanism layer integrated at the network input is responsible for feature-weighted fusion of the multi-source factors. This layer first maps the multi-source factor data at each time step into three vectors through a linear transformation: a query vector, a key vector, and a value vector.

[0031] The linear transformation employs a learnable weight matrix to project the original factor features into a high-dimensional space. The mapped vectors are then distributed to multiple attention heads for parallel computation. Each attention head independently calculates the dot product of the query vector and the key vector, and scales the result by the reciprocal of the square root of the vector dimension. Subsequently, attention weights are obtained through softmax normalization, reflecting the importance of different factors at a specific time step. The weighted summation process uses these attention weights to linearly combine the value vectors, generating the output of each attention head. Finally, the output vectors of all attention heads are concatenated into a single long vector, which is then dimensionality-reduced and integrated through another linear transformation layer to obtain the feature-weighted fusion result. This result serves as the input to a Long Short-Term Memory (LSTM) network, fed along with high-frequency component sequences for prediction. This multi-head attention mechanism captures different types of dependencies between factors through multiple sets of attention weights, enhancing the model's ability to model complex feature interactions.

[0032] The low-frequency prediction sub-model employs a multiple linear regression model, outputting the predicted values ​​of the low-frequency components. The inputs are the corresponding multi-source factor data and historical values ​​of the low-frequency component sequences. An L2 regularization term is introduced during model parameter estimation to constrain the norm of the weight vector and prevent overfitting. The L2 regularization coefficient is determined through cross-validation. The multiple linear regression model includes a bias term and multiple weight coefficients, each corresponding to the contribution of an input feature. During training, the optimal parameters are solved by minimizing the sum of the squared prediction errors and the regularization term.

[0033] The high-frequency and low-frequency prediction sub-models are configured in parallel, operating independently during training and prediction without interference. The high-frequency prediction sub-model processes the top M intrinsic mode function components with the highest center frequency, with each high-frequency component corresponding to an independent prediction instance. The low-frequency prediction sub-model processes the remaining KM low-frequency components, also with each component corresponding to a prediction instance. Corresponding prediction strategies are employed for components with different frequency characteristics. Nonlinear fluctuations in high-frequency components are captured by a long short-term memory network and an attention mechanism, while the stable trends of low-frequency components are efficiently handled by a linear regression model. The model output consists of the prediction sequences for each component, which are ultimately reconstructed into the complete load prediction value through a summation operation.

[0034] In terms of model implementation, the number of hidden layer units in the Long Short-Term Memory network is set according to the complexity of high-frequency components, and 64 units are selected. The number of heads in the multi-head attention mechanism is set to 8, and the dimension of each head is 64. The activation function of the linear transformation layer is a linear function to maintain the integrity of the features. The input features of the multiple linear regression model include numerical factors and encoded categorical factors, and the feature dimensions are consistent with the multi-source factor data.

[0035] Based on the source domain dataset, the component sequences obtained from the decomposition and the corresponding multi-source factor data are used as inputs. The combined prediction model is pre-trained with the goal of minimizing the mean square error between the predicted value and the true value of the component.

[0036] In this embodiment, the pre-training process first constructs a training sample set. Each intrinsic mode function component sequence obtained from variational mode decomposition in the source domain dataset is aligned with its corresponding multi-source factor data according to timestamps, forming an independent training sample. Each training sample contains a complete component time series and its multi-source factor feature matrix within the same time range, ensuring temporal consistency and feature integrity of the input data. The number of samples is equal to the total number K of the intrinsic mode function components obtained from the decomposition, with each component corresponding to one sample. The feature dimension of the sample is determined by the number of variables in the multi-source factor data.

[0037] The training objective is to minimize the total mean squared error (MSE) between the predicted values ​​of the components output by the combined prediction model and the true values ​​of the components. The MSE is the average of the sum of squared prediction errors for all training samples. Specifically, for each training sample, the combined prediction model generates a prediction sequence for the corresponding component based on the input multi-source factor data and the component's historical sequence, using both high-frequency and low-frequency prediction sub-models. The predicted sequence is compared point-by-point with the true sequence of the component, and the squared prediction error at each time point is calculated. The mean of the entire sequence is then taken as the MSE for that sample. The sum of the MSEs for all samples is then averaged to obtain the total MSE loss function.

[0038] The optimization process employs a backpropagation algorithm to jointly update the parameters of both the high-frequency and low-frequency prediction sub-models. Based on the gradient descent principle, the backpropagation algorithm calculates the partial derivatives of the loss function with respect to the parameters of each model layer and propagates the gradients layer-by-layer using the chain rule. The parameters of the long short-term memory network and the multi-head attention mechanism layer in the high-frequency prediction sub-model are updated synchronously, while the linear regression weights and bias terms in the low-frequency prediction sub-model are also adjusted simultaneously. The optimization algorithm selects an adaptive moment estimation optimizer, which combines a momentum term and an adaptive learning rate adjustment strategy to effectively accelerate convergence and avoid local minima. The initial learning rate is set to 0.001 and dynamically decays based on changes in loss on the validation set.

[0039] During training, the source domain dataset is divided into training and validation subsets, and an early stopping strategy is employed to prevent overfitting. After each training round, the total mean squared error on the validation set is calculated, and training terminates when the validation error no longer decreases over multiple consecutive rounds. Model parameters are initialized using the Xavier initialization method to ensure stable variance of the input and output of each layer. The entire pre-training process is executed in parallel on a graphics processing unit (GPU) to improve computational efficiency.

[0040] The pre-trained combined prediction model is transferred to the target micronet, and the model parameters are fine-tuned using sample data from the target micronet. During the fine-tuning process, the maximum mean difference between the source domain and target domain data in the hidden layer feature space of the model is calculated and used as a regularization term to form the loss function together with the prediction error.

[0041] In this embodiment, the model transfer process first applies the overall architecture and initial parameter values ​​of the combined prediction model learned on the source domain dataset during the pre-training phase to the target microgrid scenario. The sample data for the target microgrid comes from the historical operation records of the microgrid system, and the collection period is consistent with that of the source domain dataset. The sample data includes historical load sequences and multi-source factor data, and undergoes the same preprocessing procedures as the source domain, including one-hot encoding of categorical factors and normalization of numerical factors, to eliminate bias caused by differences in data distribution.

[0042] The fine-tuning process uses sample data from the target microgrid to adjust the parameters of the combined prediction model, adapting it to the specific operating characteristics of the target microgrid. During fine-tuning, the loss function consists of a prediction error term and a regularization term. The prediction error term uses mean squared error, calculating the mean squared difference between the model's predicted load and the actual load of the target microgrid, used to measure the model's prediction accuracy in the target domain. The regularization term is calculated based on the maximum mean difference between the source and target domain data in the model's hidden layer feature space; its purpose is to reduce the feature distribution differences between the source and target domains, improving the model's domain adaptability.

[0043] The maximum mean difference is calculated by measuring the distribution distance between the source and target domain data in the hidden feature space of the model using a Gaussian kernel function. Specifically, feature representations are extracted from the hidden output of the combined prediction model, and feature mappings are performed on the source and target domain samples respectively. The maximum mean difference is defined as the distance between the feature distributions of the two domains in the reproducing kernel Hilbert space, and the similarity between sample pairs is calculated using a Gaussian kernel function.

[0044] In calculating the maximum mean difference, the bandwidth parameter of the Gaussian kernel function is adaptively selected based on the dimension of the feature space. The choice of bandwidth parameter directly affects the scale sensitivity of the kernel function. In this embodiment, the square root of the feature space dimension is used as a benchmark, multiplied by a scaling factor, which is dynamically adjusted based on the distribution range of the eigenvalues. Specifically, the covariance matrix of the feature matrix is ​​first calculated, and its eigenvalue sequence is extracted. Then, the median of the eigenvalues ​​is used as a reference benchmark, combined with the feature space dimension, to calculate the bandwidth parameter. This adaptive method ensures that the bandwidth parameter can adapt to the characteristics of different feature distributions, avoiding deviations caused by manual setting.

[0045] The fine-tuning process employs gradient descent to optimize the model, simultaneously updating the parameters of both the high-frequency and low-frequency prediction sub-models in each iteration. The gradient calculation of the loss function includes the gradient of the prediction error term and the gradient of the regularization term, with the regularization gradient obtained by taking the partial derivative of the maximum mean difference with respect to the model parameters. The optimization algorithm chosen is stochastic gradient descent with momentum, using a decreasing learning rate initially set to 0.0001 and a momentum coefficient of 0.9. The iteration termination condition is based on the rate of change of loss on the validation set; fine-tuning stops when the rate of decrease of loss falls below a threshold for several consecutive iterations.

[0046] For the period to be predicted, multi-source factor data is acquired in real time and processed, then input into the fine-tuned combined prediction model to obtain the predicted values ​​of high-frequency and low-frequency components respectively; the predicted values ​​of all components are summed and reconstructed to obtain the final short-term predicted value of microgrid load.

[0047] In this embodiment, the process of acquiring multi-source factor data in real time for the period to be predicted includes synchronously collecting data from the meteorological monitoring platform, the power grid dispatch center, and the microgrid's internal monitoring system. The multi-source factor data is clearly divided into categorical factors and numerical factors. Categorical factors cover discrete variables such as weather type and time attributes, while numerical factors include environmental parameters such as temperature, humidity, and wind speed, as well as operational variables such as electricity price and distributed power generation output. The collected data is immediately preprocessed. Categorical factors are converted into binary vectors using one-hot encoding, while numerical factors are linearly scaled to the range of zero to one using the min-max normalization method. The processed multi-source factor data is aligned according to the timestamp of the period to be predicted, forming a regular input feature matrix.

[0048] Preprocessed multi-source factor data is input into a fine-tuned combined prediction model. Based on the input feature matrix, the model calls both a high-frequency prediction sub-model and a low-frequency prediction sub-model in parallel. The high-frequency prediction sub-model employs a long short-term memory network integrating a multi-head attention mechanism to predict the high-frequency intrinsic mode function (IMF) component sequences, generating predicted sequences for the high-frequency components. The low-frequency prediction sub-model uses a multiple linear regression model with L2 regularization to predict the low-frequency IMF component sequences, generating predicted sequences for the low-frequency components. The prediction processes of the two sub-models run independently to ensure the accuracy of each component prediction.

[0049] After the prediction is completed, the predicted values ​​are reconstructed. The reconstruction process sums the predicted sequences of high-frequency and low-frequency components point by point over time. That is, the predicted value at each time point is the arithmetic sum of the corresponding high-frequency and low-frequency component predicted values. The summed sequence forms a complete load prediction sequence, which represents the short-term variation trend of microgrid load. This sequence serves as the final short-term predicted value of microgrid load, used to support real-time scheduling and optimization decisions of the microgrid system.

[0050] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multi-factor high-efficiency fusion micro-grid load short-term prediction model optimization method, characterized in that, The method comprises: Collecting micro-grid load historical data and multi-source factor data, the multi-source factor data being divided into category type factors and numerical type factors; performing one-hot encoding on the category type factors, performing normalization processing on the numerical type factors, and aligning the processed multi-source factor data and the load historical data sequence according to time stamps to construct a source domain data set; Adopting a variational mode decomposition technology to decompose the aligned load historical data sequence into K intrinsic mode function components; according to the central frequency values of the components, the first M components with higher central frequency values are divided into high-frequency components, and the remaining K-M components are divided into low-frequency components; Constructing a combined prediction model, which comprises: A high-frequency prediction sub-model adopting a long short-term memory network integrated with an attention mechanism, which is used for predicting each high-frequency component sequence, wherein the attention mechanism performs feature weighted fusion on the multi-element factors input into the sub-model; A low-frequency prediction sub-model adopting a linear regression model, which is used for predicting each low-frequency component sequence; Based on the source domain data set, each component sequence and the corresponding multi-source factor data obtained by decomposition are taken as inputs, and the combined prediction model is pre-trained by taking minimizing the mean square error between the predicted value and the true value of the component as a target; The pre-trained combined prediction model is migrated to a target micro-grid, and the model parameters are fine-tuned using sample data of the target micro-grid; in the fine-tuning process, the maximum mean difference between the source domain and the target domain data in the model hidden layer feature space is calculated and taken as a regularization term together with the prediction error to form a loss function; For a to-be-predicted period, real-time multi-source factor data is obtained and processed, and then input into the fine-tuned combined prediction model to obtain the predicted values of the high-frequency components and the low-frequency components; the predicted values of all components are summed to reconstruct the final short-term micro-grid load prediction value.

2. The multi-factor high-efficiency fused microgrid load short-term prediction model optimization method of claim 1, wherein, The process of decomposition by the variational mode decomposition technology comprises: The number K of intrinsic mode function components is initially determined by calculating the power spectrum of the load historical data sequence and identifying the number of significant spectral peaks in the power spectrum; The load historical data sequence is adaptively decomposed into K narrow-band intrinsic mode function components by solving a constrained variational problem and introducing a quadratic penalty factor and a Lagrange multiplier for iterative optimization by using the variational mode decomposition technology.

3. The multi-factor high-efficiency fused microgrid load short-term prediction model optimization method of claim 2, wherein, The identification process of the significant spectral peak is: After smoothing the power spectrum, local maximum points are searched in the frequency domain, and a dynamic threshold is set, and the local maximum points with power values higher than the dynamic threshold are determined as significant spectral peaks.

4. The multi-factor high-efficiency fused microgrid load short-term prediction model optimization method of claim 1, wherein, The high-frequency prediction sub-model adopts a long short-term memory network, and a multi-head attention mechanism layer is integrated at the input end thereof, which is used for feature weighted fusion on the input multi-source factors; The low-frequency prediction sub-model adopts a multivariate linear regression model containing L2 regularization; The high-frequency prediction sub-model and the low-frequency prediction sub-model are arranged in parallel and are respectively used for processing high-frequency intrinsic mode function components and low-frequency intrinsic mode function components.

5. The multi-factor high-efficiency fused microgrid load short-term prediction model optimization method of claim 4, wherein, The multi-head attention mechanism layer specifically comprises: The multi-source factor data at each time step is respectively mapped into a query vector, a key vector and a value vector through linear transformation; The vector is assigned to multiple attention heads for parallel calculation, each attention head calculates the dot product of the query vector and the key vector, scales, obtains the attention weight through the Softmax function, and then performs weighted summation on the value vector; Finally, the outputs of all attention heads are spliced and linearly transformed to obtain the result of feature weighted fusion.

6. The multi-factor high-efficiency fused microgrid load short-term prediction model optimization method of claim 1, wherein, The pre-training process comprises: Each intrinsic modal function component sequence obtained by decomposing the source domain data set and the corresponding multi-source factor data form a training sample; The parameters of the high-frequency prediction sub-model and the low-frequency prediction sub-model are jointly optimized by using the back propagation algorithm to minimize the total mean square error between the component prediction value output by the combined prediction model and the component true value.

7. The multi-factor high-efficiency fused microgrid load short-term prediction model optimization method of claim 1, wherein, The fine-tuning process comprises: The pre-trained combined prediction model is adjusted in parameters using sample data of the target microgrid; During fine-tuning, the loss function is composed of a prediction error and a regularization term, wherein the regularization term is calculated based on the maximum mean difference in the model hidden layer feature space between the source domain and the target domain data, and the distance of the feature distribution is measured by a Gaussian kernel function.

8. The multi-factor high-efficiency fused microgrid load short-term prediction model optimization method of claim 7, wherein, In the calculation of the maximum mean difference, the bandwidth parameter of the Gaussian kernel function is adaptively selected according to the dimension of the feature space.

9. The multi-factor high-efficiency fused microgrid load short-term prediction model optimization method of claim 1, wherein, The process of obtaining the short-term load prediction value of the microgrid comprises: For the to-be-predicted period, the multi-source factor data is collected in real time, and after one-hot encoding or normalization processing, the fine-tuned combined prediction model is input; The model outputs the prediction sequence of the high-frequency component and the low-frequency component respectively, wherein the prediction value of the high-frequency component is generated by the high-frequency prediction sub-model, and the prediction value of the low-frequency component is generated by the low-frequency prediction sub-model; The prediction values of all components are summed point by point according to the time point to reconstruct a complete load prediction sequence as the short-term load prediction value of the microgrid.

Images