Power load prediction model training method and related device
By optimizing the ICEEMDAN parameters using an improved Harris Eagle optimization algorithm and combining it with the iTransformer model, the problems of parameter dependence on empirical settings and insufficient adaptive capability in existing power load forecasting technologies are solved, achieving high-precision, low-resource-consumption short-term power load forecasting.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN THERMAL POWER RES INST CO LTD
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-19
AI Technical Summary
Existing power load forecasting technologies suffer from problems such as ICEEMDAN parameters relying on empirical settings, insufficient adaptability, complex forecasting model structure, high training costs, lack of collaborative optimization in decomposition and modeling processes, and poor robustness to external feature interference.
An improved Harris Eagle Optimization (IHHO) algorithm is used to optimize the noise weights and decomposition times of ICEEMDAN. Combined with the iTransformer model, power load prediction is performed through a multi-channel input matrix and a self-attention mechanism. Relevant exogenous variables are screened, a unified multi-channel input matrix is constructed, and the model is trained.
It improves forecast accuracy and robustness, reduces computing resource requirements, and achieves high-precision, high-stability short-term power load forecasting. It is highly adaptable and can operate stably in various scenarios.
Smart Images
Figure CN122066031A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power load forecasting technology, and specifically relates to a power load forecasting model training method and related apparatus. Background Technology
[0002] Power load forecasting technology is mainly used to estimate the electricity demand of the power system over a certain period of time in the future, so as to support the dispatch of power generation resources, assist demand response, and power grid operation decisions. Among them, according to the forecast period, it can be divided into long-term (such as six months to many years), medium-term (such as monthly to quarterly), and short-term (daily to hourly) load forecasting. Furthermore, short-term power load forecasting has become a hot topic in research and application because it has the greatest impact on the real-time dispatch and safe and stable operation of the power grid.
[0003] Currently, existing power load forecasting technologies can be broadly categorized into three types: statistical methods, machine learning methods, and deep learning methods. Statistical methods include autoregression, moving average, autoregressive moving average, and various exponential smoothing models. Their advantages lie in their simple model structure and high computational efficiency, but they are less adaptable to nonlinearity and abrupt trends. Machine learning methods (such as support vector regression, random forests, and extreme gradient boosting) improve prediction accuracy by constructing multidimensional feature maps and ensemble strategies, but they rely on feature engineering and manual parameter tuning. Deep learning methods (representative models include long short-term memory networks, convolutional neural networks, and various reinforcement learning and attention mechanism models) can automatically capture complex nonlinear relationships and long-range dependencies in sequences, but they are sensitive to input data preprocessing and network hyperparameter settings, and their computational cost is relatively high.
[0004] In existing deep learning methods, to further improve prediction accuracy and robustness, many studies employ signal decomposition techniques (such as empirical mode decomposition and its improved algorithms) or integrate multiple models to extract multi-scale features and reduce noise interference. While these existing technical solutions have achieved certain results, they still suffer from drawbacks such as insufficiently intelligent parameter tuning, heavy model structures, fragmented optimization methods, and strong dependence on external variables. Consequently, they struggle to meet the comprehensive requirements of high accuracy, high stability, and high generalization ability in practical applications. Summary of the Invention
[0005] The purpose of this invention is to provide a method and related apparatus for training a power load forecasting model to solve one or more of the aforementioned technical problems. Specifically, the technical solution disclosed in this invention is a short-term power load forecasting scheme based on a combination of the Improved Harris Hawks Optimization Algorithm (IHHO), Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN), and an inverted Transformer (iTransformer) modeling structure. The trained model solves the technical problems remaining in existing solutions, such as the reliance on empirically set ICEEMDAN parameters, insufficient adaptive capability, complex forecasting model structure, high training cost, lack of collaborative optimization mechanism in the decomposition and modeling process, and poor robustness to external feature interference.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for training a power load prediction model, comprising the following steps: Obtain historical power load sequences; An improved complete adaptive noise ensemble empirical mode decomposition method, ICEEMDAN, is used to decompose historical power load sequences into multiple sub-sequences. During signal decomposition, the noise weights and decomposition times of ICEEMDAN are determined using the optimal parameter combination (Nstd) found by the improved Harris Eagle optimization algorithm (IHHO). * NE * To maximize the effect of the decomposition algorithm, Nstd * For optimized noise weights, NE * To optimize the number of decompositions; multiple subsequences include several intrinsic mode functions and one residual component; The obtained subsequences are concatenated sequentially with several selected exogenous variables in the time dimension to construct a unified multi-channel input matrix, which is then used as training samples. The iTransformer model is trained using the obtained training samples to obtain a trained power load prediction model.
[0007] A further improvement to the technical solution of this invention lies in that IHHO finds the optimal parameter combination (Nstd). * NE * The steps include: The search space is constructed using the mean absolute percentage error (MAPE) of the iTransformer model on the validation set as the fitness function. The population is initialized with settings for population size, maximum number of iterations, and linearly decreasing escape energy. In each iteration, soft encirclement and ambush behaviors of IHHO are used to update individual positions, and parameter constraints are applied. Each set of candidate parameters undergoes ICEEMDAN decomposition, and the decomposition results are used in subsequent prediction processes. The prediction error is calculated as the fitness value. After iteration, the optimal parameter combination (Nstd) is output. * NE * ).
[0008] A further improvement of the technical solution of the present invention is that, in the process of screening several exogenous variables, the Pearson correlation coefficient between the historical power load sequence and the characteristics of all candidate exogenous variables is calculated to obtain the correlation analysis results; based on the correlation analysis results, several exogenous variables that meet the set requirements for correlation with power load are screened out.
[0009] A further improvement to the technical solution of this invention is that the exogenous variables specifically include electricity price, temperature, and humidity.
[0010] A further improvement of the technical solution of the present invention lies in the step of sequentially concatenating the obtained multiple subsequences with several selected exogenous variables in the time dimension to construct a unified multi-channel input matrix and use it as a training sample. For each training sample, a time segment is constructed according to a fixed sliding window, and finally an input tensor of shape [B, W, C] is formed; where B is the batch size, W is the time step, and C is the number of channels.
[0011] A further improvement of the technical solution of the present invention is that, in the application process of the trained power load prediction model, the time series data to be predicted is first obtained based on several selected exogenous variables, and then the trained power load prediction model is used to make predictions to obtain power load prediction data.
[0012] A second aspect of the present invention provides a power load prediction model training system, comprising: The data acquisition unit is used to acquire historical power load sequences; The signal decomposition unit is used to decompose historical power load sequences using the improved Complete Adaptive Noise Integrated Empirical Mode Decomposition (ICEEMDAN) method to obtain multiple sub-sequences. During signal decomposition, the noise weights and decomposition times of ICEEMDAN are determined using the optimal parameter combination (Nstd) found by the improved Harris Eagle Optimization Algorithm (IHHO). * NE * To maximize the effect of the decomposition algorithm, Nstd * For optimized noise weights, NE *To optimize the number of decompositions; multiple subsequences include several intrinsic mode functions and one residual component; The sample acquisition unit is used to sequentially concatenate multiple subsequences with several selected exogenous variables in the time dimension to construct a unified multi-channel input matrix and use it as training samples. The training unit is used to train the iTransformer model using the obtained training samples to obtain a trained power load prediction model.
[0013] In a third aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the power load prediction model training method as described in any one of the first aspects of the present invention.
[0014] In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the power load prediction model training method as described in any one of the first aspects of the present invention.
[0015] In a fifth aspect, the present invention provides a computer program product comprising computer instructions which, when executed by a processor, implement the steps of the power load prediction model training method as described in any one of the first aspects of the present invention.
[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a method for training a power load forecasting model, specifically a short-term power load forecasting scheme based on an improved Harris Eagle optimization algorithm, ICEEMDAN signal decomposition, and iTransformer modeling structure. This invention constructs a forecasting framework that integrates parameter optimization capabilities, multi-scale feature representation capabilities, and attention modeling capabilities, providing a more rational, stable, and adaptable technical approach for refined short-term power load forecasting. Specifically, existing technologies treat ICEEMDAN signal decomposition and iTransformer as independent steps, with parameters optimized separately. Furthermore, existing Harris Eagle optimization algorithms cannot effectively optimize the ICEEMDAN parameters in the above combination. In view of the above, this invention adopts an improved Harris Eagle optimization algorithm, which improves the energy decay factor and escape energy model in the original algorithm. This improvement improves the balance between global search and local exploitation, enhances the ability to escape local optima, and makes it more suitable for parameter optimization of the model combination in this invention. This solves the problems of ICEEMDAN parameter dependence on empirical settings and insufficient adaptive capability in existing solutions. Simultaneously, it achieves an optimal combination of parameters, improving forecast accuracy. To further explain, in response to the increased difficulty in prediction due to the volatility of historical power load, this invention employs ICEEMDAN signal decomposition to decompose the original power load sequence into multiple intrinsic mode functions (IMFs). Each mode represents a specific frequency component or trend information in the data, which helps to capture the changing patterns of the data at multiple levels, enabling iTransformer to more accurately understand the different levels of data characteristics. Compared to other models, iTransformer's unique inverted attention mechanism can better capture and utilize the complex interactions between the various modal components (IMFs) decomposed by ICEEMDAN. Specifically, iTransformer innovatively treats each IMF component as an independent "variable" or "channel" and computes self-attention on these "variable" dimensions. This design enables it to efficiently learn how different frequency components (such as high-frequency noise, periodic fluctuations, and long-term trends) are interconnected and synergistic, rather than treating them as isolated information. Therefore, iTransformer can gain a deeper understanding of the multi-level features of the data revealed by ICEEMDAN and integrate all component information from a global perspective, thereby significantly improving the accuracy and robustness of predictions. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0018] Figure 1 This is a flowchart illustrating a method for training a power load prediction model in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the principle of the IHHO-ICEEMDAN-iTransformer framework in an embodiment of the present invention. Figure 3 This is a schematic diagram of dataset visualization in an embodiment of the present invention; Figure 4 This is a schematic diagram of the decomposition results of power load data in an embodiment of the present invention; Figure 5 This is a schematic diagram of the local prediction results of the experimental model in the dataset in the first prediction embodiment of the present invention; Figure 6 This is a schematic diagram of the local prediction results of the experimental model in the dataset in the second prediction embodiment of the present invention; Figure 7 This is a schematic diagram of the local prediction results of the experimental model in the dataset in the third prediction embodiment of the present invention; Figure 8 This is a schematic diagram of the local prediction results of the experimental model in the dataset in the fourth prediction embodiment of the present invention; Figure 9 This is a schematic diagram of the ICEEMDAN-DCN-Transformer architecture in existing technologies; Figure 10 This is a schematic diagram of a power load prediction model training system in an embodiment of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention; obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0020] Based on the technical solutions disclosed in the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices.
[0021] Please see Figures 1 to 4 The present invention provides a method for training a power load prediction model, comprising the following steps: Step 1: Obtain the original power load sequence.
[0022] In further optional technical solutions, if the acquired raw data contains some outliers and missing values, preprocessing is required. Specifically, for missing values, the average of the three days before and after the missing value at the same time point can be used for filling; additionally, for outliers, preset principles can be used for filtering and further processing of missing values.
[0023] In a specific embodiment of this invention, the experimental dataset is sourced from publicly available data on the Singapore Energy Market Corporation website. The selected data covers historical data from two full years, 2019 and 2020, specifically including electricity load, electricity price, and meteorological data, with a time resolution of 30 minutes, capable of reflecting the temporal characteristics of electricity load and changes in the external environment with relatively high precision. Furthermore, the experimental data is divided into three parts in a 3:1:1 ratio: training data, validation data, and test data. Additionally, considering the strong short-term dynamic characteristics of electricity load forecasting, the experimental model in this specific embodiment is set with an input sequence length of 96 and an output sequence length of 48 to achieve effective prediction of the load level for the next day. Figure 3 As shown, it illustrates the details of a portion of the samples in the dataset, which contains a total of 1440 records.
[0024] Step 2: Input the original power load sequence obtained in Step 1 into ICEEMDAN for signal decomposition. The decomposition result is several intrinsic mode functions (IMF1, IMF2, ..., IMF1). m And a residual component Res, totaling m+1 subsequences. For example... Figure 4 As shown, it illustrates the results of the training set after decomposition. It can be seen that the original non-stationary time series was successfully decomposed into several IMFs with different frequency characteristics and residual terms.
[0025] In a specific embodiment of the present invention, the specific process of signal decomposition is as follows: First, the noise weights (Nstd) and number of decompositions (NE) of ICEEMDAN are optimized using IHHO to maximize the decomposition algorithm's performance. The search space is constructed using the mean absolute percentage error (MAPE) of the prediction modules on the validation set as the fitness function. The population is initialized with settings for population size, maximum number of iterations, and linearly decreasing escape energy. In each iteration, IHHO's soft-encirclement and surprise attack behaviors are used to update individual positions, and parameter constraints are applied. Each set of candidate parameters undergoes ICEEMDAN decomposition, and its output is used in subsequent load forecasting processes, with the prediction error calculated as the fitness value. After iteration, the optimal parameter combination (Nstd) is output. * NE * ).
[0026] Then, use the optimized parameters (Nstd) * NE * The signal is decomposed; the decomposition result is several intrinsic mode functions IMF1, IMF2, ..., IMF m And a residual component Res, totaling m+1 subsequences.
[0027] Step 3, decompose all m+1 subsequences (IMF1~IMF1) obtained from ICEEMDAN. m +Res), and several exogenous variables (such as electricity price, temperature, humidity) are sequentially concatenated in the time dimension to construct a unified multi-channel input matrix; among them, for each training sample, a time segment is constructed according to a fixed sliding window, and finally an input tensor of shape [B,W,C] is formed (B is the batch size, W is the time step, and C is the number of channels).
[0028] In a specific embodiment of the present invention, during the screening of several exogenous variables, the Pearson Correlation Coefficient (PCC) is calculated between the original power load sequence and all candidate exogenous features (such as electricity price and climate data); based on the correlation results, several exogenous factors (such as electricity price, temperature, and humidity) that are highly correlated with the load are selected for subsequent prediction modeling.
[0029] Step 4: Feed the training samples into the iTransformer model, which serves as the prediction module. This model, by inverting the original Transformer module, focuses on the overall nature of the variables and is suitable for multidimensional time series data with a large number of interconnected variables. Given a multidimensional time series with a time length of T and a variable dimension of D... , This represents the entire time series for each variable indexed by n. The specific prediction process is as follows: (1) Using a multilayer perceptron (MLP) Mapped to , It contains all the temporal changes of the corresponding variable over the past period and is called VaritatToken.
[0030] (2) Use multivariate attention mechanism to analyze the correlation between each Variate Token.
[0031] ; In the formula, queries, keys, and values are represented. Indicates the projection dimension.
[0032] (3) Normalize each Varitat Token to a Gaussian distribution so that the features of all variables are under a relatively uniform distribution, thus reducing the difference in measurement units.
[0033] ; ; In the formula, Presentation layer normalization operation, A tensor representing the input. Representation tensor The first one that needs to be normalized One characteristic, Index representing the feature, from 1 to , Represents the total number of features along the normalized dimension. This represents the mean on the normalized dimension. Represents the variance along the normalization dimension. Representing the The output tensor of the layer, Representing the The output tensor of the layer, This represents the self-attention mechanism function.
[0034] (4) Use a feedforward neural network to extract the intrinsic properties of the sequence and perform layer normalization.
[0035] ; In the formula, Representing the The output tensor of the layer, Representing the The output tensor of the layer, Presentation layer normalization operation, This represents a feedforward fully connected network used for characteristic nonlinear transformations.
[0036] (5) Return the predicted sequence. ; In the formula, This represents the final predicted value sequence of the model. This represents a multilayer perceptron that acts as the prediction head. Representing the The output tensor of the layer (i.e., the last layer) encoder.
[0037] Please see Figures 5 to 8 The technical solution disclosed in this invention specifically includes: Step 1: Collecting data and processing missing and outlier values. Step 2: Optimizing ICEEMDAN parameters using IHHO. Step 3: Performing signal decomposition on the optimized ICEEMDAN. Step 4: Selecting several exogenous factors highly correlated with the load for subsequent prediction modeling. Step 5: Recombining the subsequences obtained from ICEEMDAN decomposition with the selected exogenous factors. Step 6: Training the iTransformer model to predict power load. To intuitively understand the prediction performance of the IHHO-ICEEMDAN-iTransformer model, four prediction results were randomly selected from the test set, as shown in the figure. As can be seen from the figure, regardless of whether the load is stable or fluctuating, the model's prediction curve always closely follows the actual load curve, demonstrating excellent fitting ability.
[0038] Comparative analysis of the embodiments of the present invention shows that the IHHO-ICEEMDAN-iTransformer model has higher accuracy. Each module in this combined framework plays a role in improving prediction accuracy, and its performance is the best among the compared models. The results are shown in Table 1.
[0039] Table 1. Prediction results of different models
[0040] Please see Figure 9In existing technologies, short-term power load forecasting methods based on ICEEMDAN-DCN-Transformer include: firstly, applying an improved ICEEMDAN algorithm to the original power load sequence to decompose the signal into multiple Intrinsic Mode Function (IMF) components and residuals to extract multi-scale features of different frequency components; subsequently, constructing multi-channel inputs in parallel with exogenous features such as meteorological information and holiday markers, and feeding them into a Deep Cross Network (DCN) and a standard Transformer model respectively. The DCN captures the local temporal patterns of each channel, and the global dependencies between channels are fused through the Transformer's multi-head self-attention mechanism. Finally, the fusion layer outputs a 240-hour short-term load forecast result. For experimental verification, the research team selected multi-year power load and corresponding meteorological data of Quanzhou City and compared the prediction performance of ICEEMDAN-DCN-Transformer with various single models (including CNN, LSTM, traditional Transformer, etc.) and other combined models. The results show that, under the same training and test set partitioning conditions, this method reduces key error indicators such as RMSE and MAPE by an average of approximately 6% to 8%, significantly improving prediction accuracy and enhancing the model's robustness to complex environmental factors such as sudden temperature changes and holidays. Overall, the ICEEMDAN-DCN-Transformer method, by combining multi-scale signal decomposition with parallel feature extraction and attention fusion, provides an efficient and robust technical approach for short-term power load forecasting. While retaining the advantages of traditional deep learning models, this scheme effectively removes noise interference using ICEEMDAN decomposition technology and achieves comprehensive mining and accurate prediction of load time-series characteristics through the synergistic effect of DCN and Transformer.
[0041] While the existing ICEEMDAN-DCN-Transformer short-term power load forecasting technology has certain advantages in combining multi-scale decomposition and deep learning, its parameter tuning methods are not intelligent enough, the model structure is heavy, the optimization methods are fragmented, and it is highly dependent on external variables, making it difficult to meet the comprehensive requirements of high accuracy, high stability, and high generalization ability in practical applications. Specifically: First, the parameters used in the ICEEMDAN decomposition stage of this existing technology (such as white noise intensity and number of iterations) rely on manual experience or fixed search strategies, making it difficult to achieve adaptive adjustment under different data scenarios, resulting in certain performance instability and insufficient adaptability; unreasonable parameter settings may lead to modal aliasing or information loss, thus affecting the subsequent modeling effect. Second, although the existing technology adopts a combined structure of DCN and Transformer, which can extract temporal features and global dependencies, the overall model structure is complex, consumes a lot of computing resources, has a long training time, and has high requirements for hardware configuration and parallel computing capabilities, limiting its rapid deployment and application in actual power systems. Furthermore, the model does not achieve unified optimization of the ICEEMDAN decomposition and prediction processes. These two processes are sequential and lack a global collaborative tuning mechanism, making it impossible to improve the end-to-end performance of the prediction results. Finally, when the input contains low-quality or missing exogenous features (such as meteorological factors or holiday information), the model may introduce invalid information channels, leading to misleading attention mechanisms, decreased prediction accuracy, and insufficient robustness.
[0042] In view of the above, this invention proposes a short-term power load forecasting scheme based on an improved Harris Eagle Optimization Algorithm (IHHO), ICEEMDAN signal decomposition, and iTransformer modeling structure. The improved Harris Eagle Optimization Algorithm (IHHO) adaptively optimizes the ICEEMDAN parameters, effectively avoiding decomposition errors and mode aliasing caused by empirical settings. This significantly improves the physical interpretability and frequency band separation of the obtained Intrinsic Mode Function (IMF) components. This invention organically combines decomposition parameter optimization with a multi-channel attention prediction model, realizing a closed-loop process of "parameter-decomposition-prediction," allowing the decomposition quality to directly serve the prediction effect and promoting a synergistic improvement in the overall model performance. Furthermore, this invention uses the Pearson correlation coefficient (PCC) to screen exogenous factors, introducing only meteorological and market characteristics highly correlated with the load, thereby avoiding interference from irrelevant noise and improving the robustness of the prediction model in complex environments.
[0043] Compared to multi-model ensemble schemes, the technical solution of this invention uses only a single iTransformer architecture, reducing model complexity. Simultaneously, the IHHO optimization reduces repetitive parameter tuning between decomposition and prediction, significantly shortening training time and reducing computational resource requirements. The technical solution of this invention exhibits good adaptability to power load patterns in different regions and seasons, can operate stably in various scenarios, and can be extended to longer prediction periods or the fusion of more exogenous factors according to actual needs.
[0044] In the technical solution disclosed in this invention, the IHHO algorithm is used to automatically optimize the ICEEMDAN decomposition parameters. The optimized ICEEMDAN decomposition IMF components and the exogenous features selected by PCC are then input into the iTransformer model for multi-channel attention prediction. In ICEEMDAN signal decomposition, short-term load prediction error (such as MAPE) is used as the fitness factor to automatically optimize the noise amplitude coefficient Nstd and the iteration number NE, improving decomposition quality and avoiding mode aliasing. Based on the PCC method, key factors highly correlated with load, such as electricity price, temperature, and humidity, are selected from meteorological and market data. These exogenous features are synchronously concatenated with the ICEEMDAN intrinsic mode function (IMF) channels obtained after IHHO optimization to construct a multi-channel input matrix, achieving effective coupling of endogenous features and exogenous information. In the Transformer encoder-decoder framework, frequency band adaptive embedding and channel type encoding are applied to different channels. Multi-head self-attention is used to capture complex dependencies across channels and time periods, ultimately outputting the predictions of each channel and superimposing them for reconstruction, obtaining high-precision short-term load prediction results.
[0045] The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not disclosed in the apparatus embodiments, please refer to the embodiments of the method of the present invention.
[0046] Please see Figure 10 In this embodiment of the invention, a power load prediction model training system is provided, comprising: The data acquisition unit is used to acquire historical power load sequences; The signal decomposition unit is used to decompose historical power load sequences using the improved Complete Adaptive Noise Integrated Empirical Mode Decomposition (ICEEMDAN) method to obtain multiple sub-sequences. During signal decomposition, the noise weights and decomposition times of ICEEMDAN are determined using the optimal parameter combination (Nstd) found by the improved Harris Eagle Optimization Algorithm (IHHO). * NE * To maximize the effect of the decomposition algorithm, Nstd * For optimized noise weights, NE *To optimize the number of decompositions; multiple subsequences include several intrinsic mode functions and one residual component; The sample acquisition unit is used to sequentially concatenate multiple subsequences with several selected exogenous variables in the time dimension to construct a unified multi-channel input matrix and use it as training samples. The training unit is used to train the iTransformer model using the obtained training samples to obtain a trained power load prediction model.
[0047] In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used to execute the operation of a power load prediction model training method.
[0048] In one embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM (Random Access Memory) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the power load prediction model training method in the above embodiments.
[0049] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.
[0050] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0051] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0052] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for training a power load forecasting model, characterized in that, Includes the following steps: Obtain historical power load sequences; An improved complete adaptive noise ensemble empirical mode decomposition method, ICEEMDAN, is used to decompose historical power load sequences into multiple sub-sequences. During signal decomposition, the noise weights and decomposition times of ICEEMDAN are determined using the optimal parameter combination (Nstd) found by the improved Harris Eagle optimization algorithm (IHHO). * NE * To maximize the effect of the decomposition algorithm, Nstd * For optimized noise weights, NE * To optimize the number of decompositions; multiple subsequences include several intrinsic mode functions and one residual component; The obtained subsequences are concatenated sequentially with several selected exogenous variables in the time dimension to construct a unified multi-channel input matrix, which is then used as training samples. The iTransformer model is trained using the obtained training samples to obtain a trained power load prediction model.
2. The power load forecasting model training method according to claim 1, characterized in that, IHHO finds the optimal parameter combination (Nstd) * NE * The steps include: The search space is constructed using the mean absolute percentage error (MAPE) of the iTransformer model on the validation set as the fitness function. The population is initialized with settings for population size, maximum number of iterations, and linearly decreasing escape energy. In each iteration, soft encirclement and ambush behaviors of IHHO are used to update individual positions, and parameter constraints are applied. Each set of candidate parameters undergoes ICEEMDAN decomposition, and the decomposition results are used in subsequent prediction processes. The prediction error is calculated as the fitness value. After iteration, the optimal parameter combination (Nstd) is output. * NE * ).
3. The power load forecasting model training method according to claim 1, characterized in that, In the process of screening several exogenous variables, the Pearson correlation coefficient between the historical power load sequence and the characteristics of all candidate exogenous variables is calculated to obtain the correlation analysis results; based on the correlation analysis results, several exogenous variables that meet the set requirements for correlation with power load are selected.
4. The power load forecasting model training method according to claim 3, characterized in that, Exogenous variables specifically include electricity price, temperature, and humidity.
5. The power load forecasting model training method according to claim 1, characterized in that, In the step of sequentially concatenating multiple obtained subsequences with several selected exogenous variables in the time dimension to construct a unified multi-channel input matrix and use it as training samples, for each training sample, a time segment is constructed according to a fixed sliding window, ultimately forming an input tensor of shape [B, W, C]; where B is the batch size, W is the time step, and C is the number of channels.
6. The power load forecasting model training method according to claim 1, characterized in that, In the application of a trained power load forecasting model, the time series data to be predicted is first obtained based on several selected exogenous variables, and then the trained power load forecasting model is used to make predictions to obtain power load forecast data.
7. A power load forecasting model training system, characterized in that, include: The data acquisition unit is used to acquire historical power load sequences; The signal decomposition unit is used to decompose historical power load sequences using the improved Complete Adaptive Noise Integrated Empirical Mode Decomposition (ICEEMDAN) method to obtain multiple sub-sequences. During signal decomposition, the noise weights and decomposition times of ICEEMDAN are determined using the optimal parameter combination (Nstd) found by the improved Harris Eagle Optimization Algorithm (IHHO). * NE * To maximize the effect of the decomposition algorithm, Nstd * For optimized noise weights, NE * To optimize the number of decompositions; multiple subsequences include several intrinsic mode functions and one residual component; The sample acquisition unit is used to sequentially concatenate multiple subsequences with several selected exogenous variables in the time dimension to construct a unified multi-channel input matrix and use it as training samples. The training unit is used to train the iTransformer model using the obtained training samples to obtain a trained power load prediction model.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the power load prediction model training method as described in any one of claims 1 to 6.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the power load prediction model training method as described in any one of claims 1 to 6.
10. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the power load forecasting model training method as described in any one of claims 1 to 6.