Load prediction method based on dual feature processing and error correction
The load forecasting method using variational mode decomposition and phase space reconstruction solves the problems of insufficient feature extraction and inadequate error correction in existing technologies, achieving high-precision and stable load forecasting results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA HUADIAN ENG CO LTD
- Filing Date
- 2025-12-19
- Publication Date
- 2026-05-12
AI Technical Summary
Existing load forecasting methods struggle to fully and effectively extract and utilize the complex characteristics of heating loads, are prone to feature redundancy and overfitting, and lack error correction mechanisms, resulting in poor forecast accuracy.
The variational mode decomposition algorithm is used to decompose historical load data into multiple modal components, classify them into different frequency levels according to the sample entropy, and generate initial prediction results through phase space reconstruction and load prediction sub-model, and combine them with residual prediction model for error compensation.
It significantly improves prediction accuracy and robustness, avoids confusion of multi-scale dynamic features by a single model, achieves closed-loop correction of systematic errors, and improves the long-term stability and adaptability of the model.
Smart Images
Figure CN122021995A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine learning technology, and specifically to a load prediction method based on dual feature processing and error correction. Background Technology
[0002] To improve the energy efficiency of energy systems, accurate and reliable forecasting of heat load is necessary. Because integrated energy systems involve the coupling of different energy flows, heat load often exhibits complex variation characteristics, posing a significant challenge to the optimal operation of the system.
[0003] In related technologies, time series analysis, regression analysis, support vector machines, and artificial neural networks are used for load forecasting. However, these methods are difficult to fully and effectively extract and utilize the complex feature information in the heating load, or are prone to problems such as feature redundancy and overfitting. Furthermore, they lack error correction mechanisms, resulting in poor forecast accuracy. Summary of the Invention
[0004] This invention provides a load forecasting method based on dual feature processing and error correction to solve the problem of poor accuracy in load forecasting in related technologies.
[0005] In a first aspect, the present invention provides a load forecasting method based on dual feature processing and error correction, the method comprising: Acquire multi-source time-series data of the target area; the multi-source time-series data includes historical load data and corresponding external influencing factor data; The historical load data is decomposed into multiple modal components using a variational mode decomposition algorithm; Calculate the sample entropy of each modal component separately, and classify each modal component into different frequency levels according to the magnitude of the sample entropy; Phase space reconstruction is performed based on the modal components of each frequency level and the corresponding external influencing factors data to generate a multivariate phase space dataset for each frequency level; The multivariate phase space datasets for each frequency level are input into the corresponding load forecasting sub-model to generate the forecast outputs for each frequency level and then superimpose them to obtain the initial load forecast results. A residual sequence is constructed based on the historical load data and the initial load forecast results; The residual sequence is input into the residual prediction model to obtain the load residual prediction value; The initial load forecast result is obtained by compensating the load residual forecast result with the load residual forecast value.
[0006] In one optional implementation, the step of using a variational mode decomposition algorithm to decompose the historical load data into multiple modal components includes: The variational mode decomposition algorithm is adopted, based on the optimal penalty factor and the optimal Lagrange multiplication operator, to decompose the historical load data into modal components of the target modal number; The method further includes: During the training phase, the candidate range for the number of modalities and the range of values for the penalty factor are defined. Traverse the candidate parameter combinations consisting of any mode number from the candidate range and any penalty factor from the penalty factor value range, and perform the following variational mode decomposition process on the historical loading data of the samples in the training set: A constrained variational optimization model is established with the constraint that the sum of all modal components equals the historical load data of the sample, and with the objective of minimizing the sum of the estimated bandwidths of all modal components. By introducing a penalty factor and the Lagrange multiplication operator, the constrained variational optimization model is transformed into an augmented Lagrange function; Using the number of modes in the corresponding candidate parameter combination as the target number, initialize the center frequency and corresponding modal components of the target number; The augmented Lagrangian function is iteratively updated using the alternating direction method of the multiplication operator, and each modal component and its corresponding center frequency are optimized synchronously. When the change between modal components in two adjacent iterations is less than a preset threshold, convergence is determined, and the target number of modal components is output. Repeat the variational mode decomposition process described above to obtain the set of modal components for each candidate parameter combination; The modal component sets of each candidate parameter combination are evaluated, and the candidate parameter combination with the best evaluation result is selected to determine the optimal penalty factor and the optimal Lagrange multiplier operator.
[0007] In one optional implementation, the step of calculating the sample entropy of each modal component and classifying each modal component into different frequency levels according to the sample entropy magnitude includes: Calculate the sample entropy for each modal component separately; The sample entropy of each modal component is compared with a preset sample entropy threshold, and each modal component is divided into different frequency levels according to the comparison results.
[0008] In one optional implementation, the step of reconstructing the phase space based on the modal components of each frequency level and the corresponding external influencing factor data to generate a multivariate phase space dataset for each frequency level includes: Based on the optimal delay time and optimal embedding dimension, the phase space is reconstructed according to the modal components of each frequency level and the corresponding external influencing factor data, generating a multivariate phase space dataset for each frequency level; The method further includes: During the training phase, a set of candidate delay times is constructed; Based on the first learnable weight vector, the candidate delay times in the candidate delay time set are weighted and fused to generate delay components; A multidimensional embedding vector for the phase space is constructed based on the delay component, and each embedding dimension of the phase space is weighted according to the second learnable weight vector. A maximum embedding dimension is preset, and during training, each embedding dimension is decayed sequentially according to its contribution to the prediction task until the preset sparsity condition is met, thus obtaining an effective embedding dimension. A total loss function is constructed based on the phase space reconstruction consistency loss and the load prediction sub-model loss. With the goal of minimizing the total loss function, the parameters of the first learnable weight vector, the second learnable weight vector, and the load prediction sub-model are updated synchronously through the backpropagation algorithm. When the load prediction sub-model training converges, the optimal delay time is determined based on the final first learnable weight vector, and the optimal embedding dimension is determined based on the final second learnable weight vector.
[0009] In one optional implementation, the step of inputting the multivariate phase space dataset for each frequency level into the corresponding load forecasting sub-model to generate the forecast output for each frequency level includes: The long-term time dependence features of the multivariate phase space dataset at the corresponding frequency level are extracted by the long short-term memory network layer of the load prediction sub-model to obtain the output of the first hidden layer at each time step. The attention mechanism layer of the load prediction sub-model assigns weights to the hidden layer outputs at each time step and performs a weighted summation to obtain the first context vector. The first context vector is nonlinearly transformed by the output layer of the load forecasting sub-model to generate the forecast output result corresponding to the frequency level.
[0010] In one optional implementation, the step of inputting the residual sequence into the residual prediction model to obtain the load residual prediction value includes: The long-term time dependence features of the residual sequence are extracted by the long short-term memory network layer of the residual prediction model, and the output of the second hidden layer at each time step is generated. The second context vector is obtained by assigning weights to the output of the second hidden layer at each time step through the attention mechanism layer of the residual prediction model and performing weighted summation. The output layer of the residual prediction model performs a nonlinear transformation on the second context vector to generate load residual prediction values.
[0011] Secondly, the present invention provides a load forecasting device based on dual feature processing and error correction, the device comprising: The acquisition module is used to acquire multi-source time-series data of the target area; the multi-source time-series data includes historical load data and corresponding external influencing factor data. The mode decomposition module is used to decompose the historical load data into multiple mode components using a variational mode decomposition algorithm; The frequency level module is used to calculate the sample entropy of each modal component separately and classify each modal component into different frequency levels according to the sample entropy magnitude; The phase space reconstruction module is used to reconstruct the phase space based on the modal components of each frequency level and the corresponding external influencing factors, and generate a multivariate phase space dataset for each frequency level. The load forecasting module is used to input the multivariate phase space dataset of each frequency level into the corresponding load forecasting sub-model, generate the forecast output results of each frequency level and superimpose them to obtain the initial load forecast results; The residual sequence construction module is used to construct a residual sequence based on the historical load data and the initial load prediction results. The load residual prediction module is used to input the residual sequence into the residual prediction model to obtain the load residual prediction value; The compensation module is used to compensate the initial load prediction result with the load residual prediction value to obtain the load prediction result.
[0012] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the load forecasting method based on dual feature processing and error correction described in the first aspect or any corresponding embodiment thereof.
[0013] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the load forecasting method based on dual feature processing and error correction described in the first aspect or any corresponding embodiment thereof.
[0014] Fifthly, the present invention provides a computer program product, including computer instructions for causing a computer to execute the load forecasting method based on dual feature processing and error correction described in the first aspect or any corresponding embodiment thereof.
[0015] The technical solution provided by this invention may include the following beneficial effects: The load forecasting method based on dual feature processing and error correction provided by this invention decomposes historical load data into multiple modal components using a variational mode decomposition algorithm. Each modal component is then classified into different frequency levels based on its sample entropy, effectively separating features with different emphases corresponding to different frequency levels. Furthermore, phase space reconstruction is performed based on the modal components at different frequency levels and their corresponding external influencing factors, more accurately characterizing the coupling relationships between variables and the system's intrinsic dynamic evolution. This provides high-quality feature representations for subsequent refined forecasting, significantly improving forecast accuracy. By using different load forecasting sub-models to predict multivariate phase space datasets at different frequency levels, the problem of a single model confusing multi-scale dynamic features is avoided, improving the overall fitting ability to complex load changes and thus enhancing forecast accuracy. Finally, by compensating the initial load forecast results with the load residual forecast values, closed-loop correction of systematic errors is achieved, significantly reducing forecast bias and improving the model's robustness and long-term stability. Attached Figure Description
[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of the first process of the load forecasting method based on dual feature processing and error correction according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the second process of the load forecasting method based on dual feature processing and error correction according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a load prediction sub-model based on dual feature processing and error correction according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the third process of the load forecasting method based on dual feature processing and error correction according to an embodiment of the present invention; Figure 5 This is a structural block diagram of a load forecasting device based on dual feature processing and error correction according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.
[0020] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0021] Compared to a single heating system, the integrated energy system in the park, as a new form of energy supply, can achieve tiered utilization of energy through the synergy of multiple energy sources such as electricity, gas, cooling, and heating, meeting the diverse energy needs of users. To improve the operational efficiency of the energy system, accurate and reliable forecasting of the heat load is necessary. Because the integrated energy system involves the coupling between different energy flows, the heating load often exhibits complex variation characteristics, posing a significant challenge to the optimized operation of the system.
[0022] Various load forecasting methods exist in related technologies, ranging from traditional time series analysis and regression analysis to modern machine learning algorithms such as support vector machines and artificial neural networks. However, these methods have many shortcomings. On the one hand, in terms of feature processing, load data is often highly complex and diverse, influenced by factors such as meteorological conditions and user energy consumption habits. Traditional methods struggle to comprehensively and effectively extract and utilize these complex features, and have limited ability to mine nonlinear and non-stationary features in load data, resulting in an inability to fully capture the inherent patterns of load changes. While some machine learning-based methods have improved feature extraction capabilities to some extent, they are prone to feature redundancy and overfitting when dealing with massive, high-dimensional, and noisy load data, thus limiting the model's generalization ability and prediction accuracy. On the other hand, the error problem generated during the forecasting process urgently needs to be addressed. Due to the uncertainty of load data and the dynamic changes of external factors, forecasting errors are unavoidable, and most existing forecasting models lack error correction mechanisms, making it difficult to maintain high accuracy, thereby affecting the reliable operation of integrated energy systems and scientific decision-making.
[0023] According to an embodiment of the present invention, a load forecasting method based on dual feature processing and error correction is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.
[0024] This embodiment provides a load forecasting method based on dual feature processing and error correction, which can be used in laptops, desktop computers, servers, industrial control computers, etc. Figure 1 This is a flowchart of a load forecasting method based on dual feature processing and error correction according to an embodiment of the present invention, as shown below. Figure 1 As shown, the process includes the following steps: Step S101: Obtain multi-source time-series data of the target area.
[0025] The target area can be the region requiring load forecasting. The multi-source time-series data is time-series data, including historical load data and corresponding external influencing factor data. The historical load data can be historical heat load data, historical electricity load data, etc., for the target area. The external influencing factor data can be meteorological data, such as temperature, humidity, wind speed, and solar radiation intensity. It can also be time-specific data, such as day type (weekday / holiday), hour, time period (morning / afternoon / evening), and season. This multi-source time-series data can be obtained by importing a file containing the data or by collecting it through appropriate sensors, such as obtaining electricity load data from electricity meters or temperature data from temperature sensors.
[0026] Step S102: The historical load data is decomposed into multiple modal components using a variational mode decomposition algorithm.
[0027] Variational Mode Decomposition (VMD) is an adaptive decomposition algorithm that constructs a constrained variational model to decompose the original signal into several eigenmode functions with different center frequencies and finite bandwidths. This reveals the nonlinear and non-stationary characteristics of the signal, demonstrating significant advantages in processing highly volatile data sequences such as load data. In this embodiment, after acquiring multi-source time-series data, the variational mode decomposition algorithm decomposes the historical load data in the multi-source time-series data into multiple modal components. This can separate the complex original load sequence into several eigenmode components with clear physical meaning according to different frequency characteristics, effectively revealing the multi-scale dynamic characteristics of load changes. This facilitates subsequent refined modeling and feature extraction of each component, significantly improving the prediction model's ability to capture complex load change patterns and its overall prediction accuracy.
[0028] Step S103: Calculate the sample entropy of each modal component and classify each modal component into different frequency levels according to the sample entropy.
[0029] Sample entropy, as an indicator of information complexity in a sample set, is effective in assessing the regularity and complexity of time series. A high sample entropy indicates significant irregularity and complexity, typically corresponding to high-frequency components (such as sudden disturbances or noise). A lower entropy indicates a more stable and regular sequence, belonging to low-frequency trend terms or periodic mid-frequency components. In this embodiment, the sample entropy of each modal component is calculated and sorted by magnitude. Each modal component is then classified into different frequency levels according to a preset classification rule. For example, the sorted results are divided into three different frequency levels. This quantifies the time series complexity and regularity of each modal component, enabling intelligent classification and physical meaning analysis of the decomposed signal. It effectively distinguishes transient fluctuations, periodic behaviors, and long-term trends in load changes, providing a basis for designing differentiated modeling strategies for different frequency characteristics, thereby improving the accuracy and robustness of the overall prediction model.
[0030] Step S104: Reconstruct the phase space based on the modal components of each frequency level and the corresponding external influencing factor data to generate a multivariate phase space dataset for each frequency level.
[0031] Since modal components at different frequency levels have different variation patterns and dominant factors, reconstructing the phase space independently by frequency band enables subsequent load prediction sub-models to accurately capture the variable coupling relationships and evolution trajectories at the corresponding time scales. By introducing external influencing factor data, the ability of the phase space to represent actual physical processes is enhanced, providing a structurally reasonable and information-rich input foundation for subsequent differentiated modeling and high-precision prediction, and significantly improving the model's adaptability and prediction performance to complex load changes.
[0032] Step S105: Input the multivariate phase space dataset of each frequency level into the corresponding load prediction sub-model, generate the prediction output results of each frequency level and superimpose them to obtain the initial load prediction results.
[0033] Each load forecasting sub-model has the same structure and is pre-trained using historical multi-source time-series data samples. It possesses load forecasting capabilities and can output initial load forecast sub-results for the corresponding frequency level based on the input multivariate phase space dataset. The overall initial load forecast result is obtained by superimposing the initial load forecast sub-results from each sub-model. For load variation characteristics at different frequency levels, dedicated load forecasting sub-models are used to model and generate forecast outputs for corresponding frequency bands, achieving differentiated learning and accurate capture of different frequency levels. The forecast results from each frequency band are then superimposed to reconstruct a complete initial load forecast value. Since different frequency levels have different dynamic characteristics and influencing factors, independent modeling avoids mutual interference and improves the fitting ability of each sub-model. Furthermore, the multivariate input based on phase space reconstruction further enhances the model's ability to represent variable coupling relationships and nonlinear evolution. Finally, by integrating the outputs of each sub-model, a more accurate and stable initial forecast result is obtained, providing a high-quality starting point for subsequent error correction.
[0034] Step S106: Construct a residual sequence based on the historical load data and the initial load forecast result.
[0035] Since initial predictions are inevitably subject to errors due to limitations imposed by model structure, input features, and external disturbances, residual sequences often contain learnable patterns such as periodic fluctuations, trend shifts, or responses to external disturbances. By modeling and predicting these residual sequences, dynamic compensation for the original prediction results can be achieved, thereby effectively improving overall prediction accuracy and robustness, and enhancing the model's adaptability to complex load changes. In this embodiment, a residual sequence is constructed by calculating the difference between historical load data and the initial load prediction results to explicitly capture the systematic biases and nonlinear error components that the prediction model failed to accurately fit during training.
[0036] Step S107: Input the residual sequence into the residual prediction model to obtain the load residual prediction value.
[0037] The residual prediction model has been pre-trained using historical residual sequence samples and has the function of load residual prediction. It can mine learnable features such as nonlinear patterns, periodic fluctuations and systematic deviations hidden in the input residual sequence, thereby realizing dynamic compensation for the initial load prediction results and outputting the corresponding load residual prediction values.
[0038] Step S108: The initial load forecast result is compensated by the load residual forecast value to obtain the load forecast result.
[0039] By compensating for the initial load forecast results with the load residual forecast values, the deficiencies in the original forecast can be effectively corrected, thereby improving the accuracy of the load forecast results.
[0040] The load forecasting method based on dual feature processing and error correction provided in this embodiment decomposes historical load data into multiple modal components using a variational mode decomposition algorithm. Each modal component is then classified into different frequency levels based on its sample entropy, effectively separating features with different emphases corresponding to different frequency levels. Furthermore, phase space reconstruction is performed based on the modal components at different frequency levels and their corresponding external influencing factors. This more accurately characterizes the coupling relationships between variables and the system's intrinsic dynamic evolution, providing high-quality feature representations for subsequent refined forecasting and significantly improving forecast accuracy. By using different load forecasting sub-models to predict multivariate phase space datasets at different frequency levels, the problem of a single model confusing multi-scale dynamic features is avoided, improving the overall fitting ability to complex load changes and thus enhancing forecast accuracy. Finally, by compensating the initial load forecast results with the load residual forecast values, closed-loop correction of systematic errors is achieved, significantly reducing forecast bias and improving the model's robustness and long-term stability.
[0041] This embodiment provides a load forecasting method based on dual feature processing and error correction, which can be used in laptops, desktop computers, servers, industrial control computers, etc. Figure 2 This is a flowchart of a load forecasting method based on dual feature processing and error correction according to an embodiment of the present invention, as shown below. Figure 2 As shown, the process includes the following steps: Step S201: Perform preparatory work.
[0042] To achieve the load forecasting function, this embodiment first performs preparatory work, including obtaining the optimal penalty factor and optimal Lagrange multiplication operator for variational mode decomposition, the optimal delay time and optimal embedding dimension for phase space reconstruction, training the load forecasting sub-model to be trained, and training the residual forecasting model to be trained. For ease of explanation, the stage of performing preparatory work is referred to as the training stage.
[0043] During the training phase, the optimal penalty factor and the optimal Lagrange multiplication operator for variational mode decomposition are obtained. Specifically, the candidate range for the number of modes k and the penalty factor are first defined. The range of values for is then traversed. Next, the candidate parameter combinations (k, ) consisting of any mode number within the candidate range of mode numbers and any penalty factor within the penalty factor's value range are iterated. The following variational mode decomposition process is performed on the historical load data of the samples in the training set: First, a variational problem is constructed, with the constraint that the sum of all modal components equals the historical load data of the sample, and the objective being to minimize the sum of the estimated bandwidths of all modal components. A constrained variational optimization model is then established. For example, the variational problem is described as finding k modal functions u(t) from the original signal f(t) (t being time) such that the sum of the estimated bandwidths of each mode is minimized, under the constraint that the sum of the modal components equals the original signal.
[0044] in, For the first k A function of modal components with respect to t, Let be the amplitude of the k-th modal component as a function of t. Let be the instantaneous phase of the k-th modal component as a function of t.
[0045] The constrained variational problem can be expressed in the following form:
[0046] in, For the k-th modal component after decomposition, Let be the center frequency of the k-th modal component. This is the Dirac function.
[0047] Next, the variational problem is solved by introducing a penalty factor. and Lagrange multiplication operators ( The constrained variational optimization model is transformed into an unconstrained variational problem, yielding the augmented Lagrangian function. This penalty factor... This is the quadratic penalty factor. An example of an augmented Lagrangian function is as follows:
[0048] Then, using the number of modes in the corresponding candidate parameter combination as the target number, initialize the center frequency and corresponding modal components of the target number.
[0049] Next, the augmented Lagrangian function is iteratively updated using the alternating direction method of the multiplication operator, simultaneously optimizing each modal component and its corresponding center frequency. For example, the alternating direction method of the multiplication operator is introduced, and updates are performed through continuous alternating calculations. , , To find the optimal solution, we then use the Fourier isometry transform to obtain the following expression:
[0050]
[0051] in, For the center frequency, This refers to the k-th modal component in the (n+1)th alternating calculation; The center frequency of the k-th modal component in the (n+1)th alternating calculation; For the Lagrange multiplication operator in the (n+1)th alternating calculation, adding the "^" symbol indicates the corresponding quantity after the Fourier transform, for example... This represents the k-th modal component in the (n+1)-th alternating calculation after the Fourier transform.
[0052] Finally, when the change between modal components in two consecutive iterations is less than a preset threshold, convergence is determined, and the target number of modal components is output. For example, by solving the above formula, decomposition is completed and k modal components are output when the following condition is met:
[0053] Where b represents the number of iterations. This indicates a preset threshold, used to indicate the accuracy of the judgment.
[0054] Repeat the variational mode decomposition process described above to obtain the modal component set of each candidate parameter combination. Evaluate the modal component set of each candidate parameter combination and select the candidate parameter combination with the best evaluation result to determine the optimal penalty factor and the optimal Lagrange multiplier operator.
[0055] During the training phase, the optimal latency and optimal embedding dimension for phase space reconstruction are obtained. Model-adaptive phase space reconstruction addresses the latency issues inherent in traditional phase space reconstruction. and embedding dimension The problem of relying on human experience for confirmation and being unable to dynamically adapt to changes in data characteristics is addressed by proposing a method for co-optimization of reconstructed parameters and task models. This method integrates parameter optimization of phase space reconstruction into the model training process, and automatically learns the optimal latency time through data-driven learning. and optimal embedding dimension This allows the reconstructed phase space to retain the nonlinear dynamic characteristics of the original time-series data while also adapting to the input requirements of the load forecasting model. Specifically, a set of candidate delay times is first constructed. An example multivariate phase space dataset for each frequency level is as follows:
[0056] Where p represents the frequency level and k represents the k-th modal component.
[0057] Set candidate delay time gather:
[0058] Introducing the first learnable weight vector:
[0059] Next, based on the first learnable weight vector, the candidate delay times in the candidate delay time set are weighted and fused to generate delay components.
[0060] Then, a multidimensional embedding vector in the phase space is constructed based on the delay components. For example, the j-th component of the embedding vector is a weighted combination of the candidate delay times:
[0061] Introducing a second learnable weight vector:
[0062] The embedding vector is obtained by weighting each embedding dimension of the phase space according to the second learnable weight vector:
[0063] For embedding dimension Preset maximum embedding dimension During training, the embedding dimensions are progressively reduced in weight according to their contribution to the prediction task until a preset sparsity condition is met, resulting in the effective embedding dimension. For example, during training, the weights of dimensions that contribute nothing to the reconstruction result gradually decay to 0, and the final effective embedding dimension is:
[0064] in, This indicates the threshold for determining effective weights, which is greater than... This indicates that it contributed to the reconstruction effect.
[0065] Next, a total loss function is constructed based on the phase space reconstruction consistency loss and the load forecasting sub-model loss. For example, the total loss function consists of two parts: the phase space reconstruction consistency loss and the phase space reconstruction consistency loss. and load prediction sub-model loss That is, the total loss function is:
[0066] in, This is the balancing coefficient, used to control the weight of the two parts of the loss.
[0067] Then, with the goal of minimizing the total loss function, the parameters of the first learnable weight vector, the second learnable weight vector, and the load forecasting sub-model are synchronously updated through backpropagation. When the load forecasting sub-model training converges, the optimal delay time is determined based on the final first learnable weight vector, and the optimal embedding dimension is determined based on the final second learnable weight vector. Thus, in the subsequent inference stage, a phase space vector satisfying the optimal prediction effect is constructed using the optimal delay time and the optimal embedding dimension. By jointly training the reconstruction parameters (first learnable weight vector, second learnable weight vector) and the model parameters (parameters of the load forecasting sub-model) online, the sequence reconstruction and the load forecasting sub-model are deeply coupled, effectively solving the pain point in traditional phase space reconstruction where the delay time and embedding dimension cannot dynamically adapt to changes in data characteristics. This achieves end-to-end joint optimization, improving the accuracy and robustness of the load forecasting sub-model.
[0068] When training the load forecasting sub-model to be trained, it should be noted that the network structure of the load forecasting sub-model to be trained is the same as that of the load forecasting sub-model in the inference stage, but the network parameters are different. The collected historical multi-source time series data samples can be subjected to variational mode decomposition and phase space reconstruction to obtain multivariate phase space dataset samples. The load forecasting sub-model to be trained is trained using multivariate phase space dataset samples to optimize the network parameters of the load forecasting sub-model to be trained, thus obtaining the load forecasting sub-model.
[0069] For example, Figure 3 This is a schematic diagram of the load prediction sub-model according to an embodiment of the present invention. An LSTM-Attention load prediction sub-model is constructed, where the Long Short-Term Memory (LSTM) network layer is used to learn the long-term time-series dependencies of load data, and the Attention mechanism layer is used to assign different weights to the output states of the LSTM network layer, highlighting key features affecting the load, strengthening the model's learning of key features, and reducing the impact of redundant information. The input layer of the load prediction sub-model takes samples X(X0) from the multivariate phase space dataset. t The multivariate phase space dataset sample at time t is used as the input to the Long Short-Term Memory (LSTM) network layer. The LSTM network layer learns the sample features to obtain the output h of the corresponding hidden layer, which is then fed into the Attention layer. The Attention layer assigns different weights to the input features and updates them continuously. The weight coefficients are calculated as follows:
[0070]
[0071]
[0072] in, for t The hidden layer vector at time step 1. Let J be the hidden layer vector at time j. Here is the attention weight matrix. , This is the weight coefficient matrix. for t Hidden layer output at any given time, For the output of the attention mechanism layer, This is the bias vector.
[0073] After the fully connected layer of the load forecasting sub-model, the expression for the prediction result of the output layer is:
[0074] in, To predict the output value, This is the weight matrix. b For bias vectors, This represents the Sigmoid function.
[0075] When training the residual prediction model to be trained, it should be noted that the network structure and training process of the residual prediction model are similar to those of the load prediction sub-model. The difference is that the training data used is historical residual sequence samples and the function is to predict load residuals based on the residual sequence. This will not be elaborated here.
[0076] Step S202: Obtain multi-source time-series data of the target area.
[0077] The multi-source time-series data includes historical load data and corresponding external influencing factor data.
[0078] For example, import standardized datasets, including multi-source time-series data such as load values, temperature, humidity, and day type, and support CSV, xls, and xlsx file formats.
[0079] Optionally, data cleaning can be performed on the multi-source time series data. For example, the upper and lower quartile method and the K-nearest neighbor filling method can be used to detect outliers and fill missing values. Spearman correlation analysis can be used to screen out the factors affecting load changes. The preprocessed multi-source time series data can then be used in subsequent steps.
[0080] Step S203: The historical load data is decomposed into multiple modal components using a variational mode decomposition algorithm.
[0081] Specifically, a variational mode decomposition algorithm is used, based on the optimal penalty factor and the optimal Lagrange multiplication operator, to decompose the historical load data into modal components of the target modal number.
[0082] Step S204: Calculate the sample entropy of each modal component and classify each modal component into different frequency levels according to the sample entropy.
[0083] Specifically, the sample entropy of each modal component is calculated separately; the sample entropy of each modal component is compared with a preset sample entropy threshold, and each modal component is classified into different frequency levels according to the comparison results.
[0084] For example, for a signal The steps for calculating sample entropy are as follows: a. Reconstructing an m-dimensional vector:
[0085] b. Calculate the maximum difference between vectors:
[0086] c. Statistical similarity ratio:
[0087] d. Calculate the average and add dimensions:
[0088] e. The sample entropy value is:
[0089] For example, the frequency level includes high frequency, intermediate frequency and low frequency, and the modal components are classified into high frequency modal components, intermediate frequency modal components and low frequency modal components.
[0090] Step S205: Reconstruct the phase space based on the modal components of each frequency level and the corresponding external influencing factor data to generate a multivariate phase space dataset for each frequency level.
[0091] Specifically, based on the optimal delay time and the optimal embedding dimension, the phase space is reconstructed according to the modal components of each frequency level and the corresponding external influencing factor data, generating a multivariate phase space dataset for each frequency level.
[0092] Step S206: Input the multivariate phase space dataset of each frequency level into the corresponding load forecasting sub-model, generate the forecast output results of each frequency level and superimpose them to obtain the initial load forecast results; Specifically, the long short-term memory network layer of the load prediction sub-model extracts the long-term time-dependent features of the multivariate phase space dataset at the corresponding frequency level to obtain the first hidden layer output at each time step; the attention mechanism layer of the load prediction sub-model assigns weights to the hidden layer output at each time step and performs weighted summation to obtain the first context vector; the output layer of the load prediction sub-model performs a nonlinear transformation on the first context vector to generate the prediction output result at the corresponding frequency level.
[0093] Step S207: Construct a residual sequence based on the historical load data and the initial load forecast result.
[0094] Specifically, the difference between historical load data and the initial load forecast is calculated to construct a residual sequence.
[0095] Step S208: Input the residual sequence into the residual prediction model to obtain the load residual prediction value.
[0096] Specifically, the long-term time-dependent features of the residual sequence are extracted through the long short-term memory network layer of the residual prediction model to generate the output of the second hidden layer at each time step. The attention mechanism layer of the residual prediction model assigns weights to the output of the second hidden layer at each time step and performs a weighted summation to obtain the second context vector. The output layer of the residual prediction model then performs a nonlinear transformation on this second context vector to generate the predicted load residual value. In other words, this residual prediction model can be configured with a network structure similar to that of the load prediction sub-model. Since the residual sequence and the original load sequence (historical load data) have similar nonlinear and non-stationary characteristics in terms of time dependence (such as periodic fluctuations, trend changes, and sudden disturbance responses), using an LSTM-Attention model with the same structure ensures that the residual prediction model has equivalent time series modeling capabilities, effectively capturing complex dynamic patterns in the residuals. Furthermore, the shared network architecture helps reduce system design complexity, improves model reusability and training stability, and ensures that the attention mechanism accurately focuses on key historical error segments, thereby achieving efficient learning and high-precision compensation of the initial prediction error, further enhancing the robustness and generalization performance of the overall prediction system.
[0097] Step S209: The initial load forecast result is compensated by the load residual forecast value to obtain the load forecast result.
[0098] By compensating the initial load forecast result with the load residual forecast value of the secondary forecast, an error correction mechanism can be established to further reduce the forecast error.
[0099] As one or more specific application embodiments of the present invention, the optimal implementation scheme or the scheme that the inventors most want to embody is described in combination with the specific application scenario.
[0100] Figure 4 This is a flowchart of a load forecasting method based on dual feature processing and error correction according to an embodiment of the present invention. First, historical load data and meteorological data (e.g., temperature, humidity) are collected. Then, the historical load data (load sequence) is modally decomposed using the Variational Mode Decomposition (VMD) algorithm, and classified based on the sample entropy of the mode decomposition results to obtain high-frequency, mid-frequency, and low-frequency components. Spearman correlation analysis is then used to filter the meteorological data, quantifying the correlation between historical load data and meteorological data to obtain the filtered feature data. This filtered feature data is combined with the high-frequency, mid-frequency, and low-frequency components to obtain high-frequency component multivariate time series, mid-frequency component multivariate time series, and low-frequency component multivariate time series. These are then subjected to adaptive phase space reconstruction and input into the corresponding load forecasting sub-models. For example, the high-frequency component multivariate time series is input into the first load forecasting sub-model, which includes a first Long Short-Term Memory (LSTM) network layer and a first attention mechanism layer (…). Figure 4 (Attention is used to represent all values). The outputs of each load prediction sub-model are superimposed to obtain the initial load prediction value. The residual sequence is then input into the residual prediction model (which includes a Long Short-Term Memory (LSTM) network layer and an Attention mechanism layer) to obtain the error correction sequence. The initial load prediction value is then compensated using the error correction sequence to obtain the final load prediction value.
[0101] This embodiment also provides a load forecasting device based on dual feature processing and error correction, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.
[0102] This embodiment provides a load forecasting device based on dual feature processing and error correction, such as... Figure 5 As shown, it includes: The acquisition module 501 is used to acquire multi-source time-series data of the target area; the multi-source time-series data includes historical load data and corresponding external influencing factor data. The mode decomposition module 502 is used to decompose the historical load data into multiple modal components using a variational mode decomposition algorithm; The frequency level module 503 is used to calculate the sample entropy of each modal component separately and classify each modal component into different frequency levels according to the sample entropy magnitude. The phase space reconstruction module 504 is used to reconstruct the phase space based on the modal components of each frequency level and the corresponding external influencing factor data, and generate a multivariate phase space dataset for each frequency level. The load forecasting module 505 is used to input the multivariate phase space dataset of each frequency level into the corresponding load forecasting sub-model, generate the forecast output results of each frequency level and superimpose them to obtain the initial load forecasting results. The residual sequence construction module 506 is used to construct a residual sequence based on the historical load data and the initial load forecast result; The load residual prediction module 507 is used to input the residual sequence into the residual prediction model to obtain the load residual prediction value; The compensation module 508 is used to compensate the initial load forecast result with the load residual forecast value to obtain the load forecast result.
[0103] In an optional implementation, the mode decomposition module is further configured to: The variational mode decomposition algorithm is adopted, based on the optimal penalty factor and the optimal Lagrange multiplication operator, to decompose the historical load data into modal components with the target number of modes; The device also includes a training module for: During the training phase, the candidate range for the number of modalities and the range of values for the penalty factor are defined. Traverse the candidate parameter combinations consisting of any mode number from the candidate range and any penalty factor from the penalty factor value range, and perform the following variational mode decomposition process on the historical loading data of the samples in the training set: A constrained variational optimization model is established with the constraint that the sum of all modal components equals the historical load data of the sample, and with the objective of minimizing the sum of the estimated bandwidths of all modal components. By introducing a penalty factor and the Lagrange multiplication operator, the constrained variational optimization model is transformed into an augmented Lagrange function. Using the number of modes in the corresponding candidate parameter combination as the target number, initialize the center frequency and corresponding modal components of the target number; The augmented Lagrangian function is iteratively updated using the alternating direction method of the multiplication operator, and each modal component and its corresponding center frequency are optimized simultaneously. When the change between modal components in two adjacent iterations is less than a preset threshold, convergence is determined, and the target number of modal components is output. Repeat the variational mode decomposition process described above to obtain the set of modal components for each candidate parameter combination; The modal component sets of each candidate parameter combination are evaluated, and the candidate parameter combination with the best evaluation result is selected to determine the optimal penalty factor and the optimal Lagrange multiplier operator.
[0104] In one alternative implementation, the frequency level module is further configured to: Calculate the sample entropy for each modal component separately; The sample entropy of each modal component is compared with a preset sample entropy threshold, and each modal component is divided into different frequency levels according to the comparison results.
[0105] In an optional implementation, the phase space reconstruction module is further configured to: Based on the optimal delay time and optimal embedding dimension, the phase space is reconstructed according to the modal components of each frequency level and the corresponding external influencing factor data, generating a multivariate phase space dataset for each frequency level; This training module is also used for: During the training phase, a set of candidate delay times is constructed; Based on the first learnable weight vector, the candidate delay times in the candidate delay time set are weighted and fused to generate delay components; A multidimensional embedding vector for the phase space is constructed based on the delay component, and each embedding dimension of the phase space is weighted according to the second learnable weight vector. A maximum embedding dimension is preset, and during training, each embedding dimension is decayed sequentially according to its contribution to the prediction task until the preset sparsity condition is met, thus obtaining an effective embedding dimension. A total loss function is constructed based on the phase space reconstruction consistency loss and the load prediction sub-model loss. With the goal of minimizing the total loss function, the parameters of the first learnable weight vector, the second learnable weight vector, and the load prediction sub-model are updated synchronously through the backpropagation algorithm. When the load prediction sub-model training converges, the optimal delay time is determined based on the final first learnable weight vector, and the optimal embedding dimension is determined based on the final second learnable weight vector.
[0106] In an optional implementation, the load forecasting module is further configured to: The long-term time dependence features of the multivariate phase space dataset at the corresponding frequency level are extracted by the long short-term memory network layer of the load prediction sub-model to obtain the output of the first hidden layer at each time step. The attention mechanism layer of the load prediction sub-model assigns weights to the hidden layer outputs at each time step and performs a weighted summation to obtain the first context vector. The first context vector is nonlinearly transformed by the output layer of the load prediction sub-model to generate the prediction output result corresponding to the frequency level.
[0107] In an optional implementation, the load residual prediction module is further configured to: The long-term time dependence features of the residual sequence are extracted by the long short-term memory network layer of the residual prediction model, and the output of the second hidden layer at each time step is generated. The second context vector is obtained by assigning weights to the output of the second hidden layer at each time step through the attention mechanism layer of the residual prediction model and performing weighted summation. The second context vector is nonlinearly transformed by the output layer of the residual prediction model to generate the load residual prediction value.
[0108] The load forecasting device based on dual feature processing and error correction provided in this embodiment of the invention can execute the load forecasting method based on dual feature processing and error correction provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the above modules and units are the same as in the corresponding embodiments described above, and will not be repeated here.
[0109] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.
[0110] The following is a detailed reference. Figure 6 This diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 602 or a program loaded from memory 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via bus 604. An input / output (I / O) interface 605 is also connected to bus 604.
[0111] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.
[0112] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a memory 608, or installed from a ROM 602. When the computer program is executed by the processor 601, it performs the functions defined in the load forecasting method based on dual feature processing and error correction according to embodiments of the present invention.
[0113] Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.
[0114] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium after being downloaded via a network. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the load forecasting method based on dual feature processing and error correction shown in the above embodiments is implemented.
[0115] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.
[0116] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the invention.
Claims
1. A load forecasting method based on dual feature processing and error correction, characterized in that, The method includes: Acquire multi-source time-series data of the target area; the multi-source time-series data includes historical load data and corresponding external influencing factor data; The historical load data is decomposed into multiple modal components using a variational mode decomposition algorithm; Calculate the sample entropy of each modal component separately, and classify each modal component into different frequency levels according to the magnitude of the sample entropy; Phase space reconstruction is performed based on the modal components of each frequency level and the corresponding external influencing factors data to generate a multivariate phase space dataset for each frequency level; The multivariate phase space datasets for each frequency level are input into the corresponding load forecasting sub-model to generate the forecast outputs for each frequency level and then superimpose them to obtain the initial load forecast results. A residual sequence is constructed based on the historical load data and the initial load forecast results; The residual sequence is input into the residual prediction model to obtain the load residual prediction value; The initial load forecast result is obtained by compensating the load residual forecast result with the load residual forecast value.
2. The method according to claim 1, characterized in that, The historical load data is decomposed into multiple modal components using a variational mode decomposition algorithm, including: The variational mode decomposition algorithm is adopted, based on the optimal penalty factor and the optimal Lagrange multiplication operator, to decompose the historical load data into modal components of the target modal number; The method further includes: During the training phase, the candidate range for the number of modalities and the range of values for the penalty factor are defined. Traverse the candidate parameter combinations consisting of any mode number from the candidate range and any penalty factor from the penalty factor value range, and perform the following variational mode decomposition process on the historical loading data of the samples in the training set: A constrained variational optimization model is established with the constraint that the sum of all modal components equals the historical load data of the sample, and with the objective of minimizing the sum of the estimated bandwidths of all modal components. By introducing a penalty factor and the Lagrange multiplication operator, the constrained variational optimization model is transformed into an augmented Lagrange function; The target number is determined by taking the number of modes in the corresponding candidate parameter combination as the target number, and the center frequency and corresponding modal components of the target number are initialized. The augmented Lagrangian function is iteratively updated using the alternating direction method of the multiplication operator, and each modal component and its corresponding center frequency are optimized synchronously. When the change between modal components in two adjacent iterations is less than a preset threshold, convergence is determined, and the target number of modal components is output. Repeat the variational mode decomposition process described above to obtain the set of modal components for each candidate parameter combination; The modal component sets of each candidate parameter combination are evaluated, and the candidate parameter combination with the best evaluation result is selected to determine the optimal penalty factor and the optimal Lagrange multiplier operator.
3. The method according to claim 1, characterized in that, The step of calculating the sample entropy of each modal component and classifying each modal component into different frequency levels according to the magnitude of the sample entropy includes: Calculate the sample entropy for each modal component separately; The sample entropy of each modal component is compared with a preset sample entropy threshold, and each modal component is classified into different frequency levels according to the comparison results.
4. The method according to claim 1, characterized in that, The step involves reconstructing the phase space based on the modal components at each frequency level and the corresponding external influencing factors, generating a multivariate phase space dataset for each frequency level, including: Based on the optimal delay time and optimal embedding dimension, the phase space is reconstructed according to the modal components of each frequency level and the corresponding external influencing factor data, generating a multivariate phase space dataset for each frequency level; The method further includes: During the training phase, a set of candidate delay times is constructed; Based on the first learnable weight vector, the candidate delay times in the candidate delay time set are weighted and fused to generate delay components; A multidimensional embedding vector for the phase space is constructed based on the delay component, and each embedding dimension of the phase space is weighted according to the second learnable weight vector. A maximum embedding dimension is preset, and during training, each embedding dimension is decayed sequentially according to its contribution to the prediction task until the preset sparsity condition is met, thus obtaining an effective embedding dimension. A total loss function is constructed based on the phase space reconstruction consistency loss and the load prediction sub-model loss. With the goal of minimizing the total loss function, the parameters of the first learnable weight vector, the second learnable weight vector, and the load prediction sub-model are updated synchronously through the backpropagation algorithm. When the load prediction sub-model training converges, the optimal delay time is determined based on the final first learnable weight vector, and the optimal embedding dimension is determined based on the final second learnable weight vector.
5. The method according to any one of claims 1 to 4, characterized in that, The step of inputting the multivariate phase space dataset for each frequency level into the corresponding load forecasting sub-model to generate the forecast output for each frequency level includes: The long-term time dependence features of the multivariate phase space dataset at the corresponding frequency level are extracted by the long short-term memory network layer of the load prediction sub-model to obtain the output of the first hidden layer at each time step. The attention mechanism layer of the load prediction sub-model assigns weights to the hidden layer outputs at each time step and performs a weighted summation to obtain the first context vector. The first context vector is nonlinearly transformed by the output layer of the load forecasting sub-model to generate the forecast output result corresponding to the frequency level.
6. The method according to claim 5, characterized in that, The step of inputting the residual sequence into the residual prediction model to obtain the load residual prediction value includes: The long-term time dependence features of the residual sequence are extracted by the long short-term memory network layer of the residual prediction model, and the output of the second hidden layer at each time step is generated. The second context vector is obtained by assigning weights to the output of the second hidden layer at each time step through the attention mechanism layer of the residual prediction model and performing weighted summation. The output layer of the residual prediction model performs a nonlinear transformation on the second context vector to generate load residual prediction values.
7. A load forecasting device based on dual feature processing and error correction, characterized in that, The device includes: The acquisition module is used to acquire multi-source time-series data of the target area; the multi-source time-series data includes historical load data and corresponding external influencing factor data. The mode decomposition module is used to decompose the historical load data into multiple mode components using a variational mode decomposition algorithm; The frequency level module is used to calculate the sample entropy of each modal component separately and classify each modal component into different frequency levels according to the sample entropy magnitude; The phase space reconstruction module is used to reconstruct the phase space based on the modal components of each frequency level and the corresponding external influencing factors, and generate a multivariate phase space dataset for each frequency level. The load forecasting module is used to input the multivariate phase space dataset of each frequency level into the corresponding load forecasting sub-model, generate the forecast output results of each frequency level and superimpose them to obtain the initial load forecast results; The residual sequence construction module is used to construct a residual sequence based on the historical load data and the initial load prediction results. The load residual prediction module is used to input the residual sequence into the residual prediction model to obtain the load residual prediction value; The compensation module is used to compensate the initial load prediction result with the load residual prediction value to obtain the load prediction result.
8. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory stores computer instructions, and the processor executes the load forecasting method based on dual feature processing and error correction as described in any one of claims 1 to 6 by executing the computer instructions.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the load forecasting method based on dual feature processing and error correction as described in any one of claims 1 to 6.
10. A computer program product, characterized in that, Includes computer instructions for causing a computer to execute the load forecasting method based on dual feature processing and error correction as described in any one of claims 1 to 6.