Multivariate load power prediction method and device based on multi-scale feature extraction

Abstract

The application is a multi-element load power prediction method and device based on multi-scale feature extraction, aiming at the problems of coarse frequency band division, single feature extraction and ignoring the coupling relationship of multi-element load frequency band in the existing prediction scheme, comprising: data preprocessing of multi-element load to obtain preprocessed data; quantifying the correlation and coupling characteristics of meteorological features and multi-element load, and screening out strong correlation meteorological features; decomposing the preprocessed data, and reconstructing the sequence based on the sample entropy value of the decomposition result to obtain a reconstructed sequence; based on the coupling characteristics, using the Prophet model to extract the change rule of the multi-time scale of the preprocessed data as the multi-time scale feature; inputting the strong correlation meteorological features, the multi-time scale feature and the reconstructed sequence into the XGboost model for prediction and fitting to obtain the final prediction result. The scheme of the application can improve the power prediction accuracy and working condition robustness of the comprehensive energy system electric, heat and cold multi-element load.

Description

Technical Field

[0001] This invention relates to the field of power system load power prediction technology, and in particular to a multi-dimensional load power prediction method and apparatus based on multi-scale feature extraction. Background Technology

[0002] Integrated energy parks featuring deep interaction between power generation, grid, load, and storage, and multi-energy complementarity, are currently being widely promoted and applied. The high proportion of distributed wind and solar power connected to the grid, the continuous growth of electric vehicles, controllable flexible loads, and energy-cooling coupled devices have led to new characteristics on the load side: strong randomness, time-varying nonlinearity, and deep multi-energy coupling. Coupled with frequent extreme weather events, the uncertainty on both the power generation and load sides has significantly increased, posing challenges to the safe and stable operation of the power grid, power consumption dispatch, demand response, and market transactions. High-precision multi-dimensional load power forecasting is a crucial prerequisite for achieving power balance, optimized energy storage configuration, virtual power plant regulation, and integrated energy economic dispatch in the new power system.

[0003] Integrated energy systems simultaneously contain diverse loads, including electricity, heat, and cooling. Load timing includes low-frequency trend components dominated by seasonal and weekday rhythms, mid-frequency cyclical components influenced by weather and equipment operating cycles, and high-frequency fluctuation components caused by equipment start-up and shutdown, weather disturbances, and users' random energy consumption behaviors. The physical causes of signals in different frequency bands differ significantly. Current forecasting technologies for the diverse loads of new power systems still have significant shortcomings. First, traditional forecasting methods directly input the original multi-variable load time series into the model without frequency band decoupling. Low-frequency trends, mid-frequency cycles, and high-frequency noise are mixed together, and feature extraction is mostly limited to meteorological characteristics, lacking analysis of the fluctuation characteristics of the multi-variable load power series themselves. Under conditions of high penetration of new energy sources and extreme weather disturbances, high-frequency random disturbances can interfere with the model's learning of long-term load trends and periodic patterns, causing amplified forecasting errors and making it difficult to meet the forecasting reliability requirements of new power system dispatching.

[0004] Second, there is a strong energy coupling relationship among multiple loads, including electricity, heat, and cold. Existing technologies mostly model and predict various loads independently, simply splicing external features at the level of raw data, and lack the ability to explore the coupling relationship between different load components at the frequency band level. This results in poor robustness under typical disturbance scenarios of new power systems, such as load changes and extreme weather.

[0005] In summary, existing prediction schemes suffer from technical problems such as coarse frequency band division, simplistic feature extraction, and neglect of the coupling relationship between multiple load frequency bands. Summary of the Invention

[0006] To address the technical problems of existing prediction schemes, such as coarse frequency band division, single feature extraction, and neglect of the coupling relationship of multiple load frequency bands, this invention provides a method and device for predicting the power of multiple loads based on multi-scale feature extraction. This method improves the accuracy and robustness of power prediction for multiple loads (electricity, heat, and cooling) in integrated energy systems under new power system scenarios, and provides support for the coordinated scheduling of power generation, grid, load, and storage in new power systems.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: A multivariate load power prediction method based on multi-scale feature extraction includes the following steps: S1, perform data preprocessing on the multivariate workload by removing outliers and filling in missing values ​​to obtain preprocessed data; S2 uses Spearman rank correlation coefficient to quantify the correlation and coupling characteristics between meteorological features and multivariate loads, and screens out strongly correlated meteorological features; among which the coupling characteristics are the correlation with multivariate loads and the coupling characteristics between multivariate loads themselves; S3. The preprocessed data is decomposed using the CEEMDAN algorithm, and the sequence is reconstructed based on the sample entropy value of the decomposition result to obtain the reconstructed sequence. S4. Based on coupling features, the Prophet model is used to extract the variation patterns of preprocessed data across multiple time scales, which are then used as multi-time scale features. S5 inputs strongly correlated meteorological features, multiple time scale features, and reconstructed sequences into the XGboost model for prediction and fitting, and obtains the final prediction results.

[0008] The Spearman correlation coefficient method is used to calculate the correlation coefficient between each feature and the target power, retaining strongly correlated features and eliminating weakly correlated redundant features, thereby simplifying the input feature set and improving the quality of input features for the prediction model.

[0009] Spearman's rank correlation coefficient is used to measure the monotonic correlation between meteorological characteristics and load power, and to achieve initial characteristic selection.

[0010] Furthermore, the CEEMDAN algorithm steps are as follows: Different Gaussian white noise sequences are added to the preprocessed data to construct a pre-decomposition sequence; Perform EMD decomposition on the preprocessed data to obtain the first IMF component; The residual is calculated by subtracting the first IMF component from the preprocessed data. The residual is then decomposed into EMD to obtain the remaining IMF components. This process is repeated until the residual is a monotonic sequence from which no meaningful IMF components can be decomposed.

[0011] We introduce sample entropy (SE) as a metric for component complexity.

[0012] Furthermore, the sample entropy calculation includes: The sample entropy value is calculated for each decomposed component using sample entropy, and is used to evaluate the component complexity. Based on the sample entropy calculation results, K-Means clustering is used to reorganize each decomposed IMF component to obtain three reorganized components of load power: high-frequency random component, mid-frequency periodic component, and low-frequency trend component. These components serve as reconstructed sequences to reflect the fluctuation characteristics of load power at different time scales.

[0013] Furthermore, the Prophet model is used to extract potential multi-timescale fluctuation characteristics in load power. Specifically, the time series is divided into four parts using the Prophet model: trend component, periodic component, holiday effect component, and noise component. The trend component represents the annual variation of load and is calculated using a nonlinear logistic regression function. The periodic component represents the periodic characteristics of power load at different time scales and is constructed using Fourier series. The holiday effect component represents the impact of holidays or sudden events on the series fluctuations and is modeled using a matrix that includes holidays and special events. The noise component represents unpredictable fluctuations and is represented using a normal distribution function.

[0014] Furthermore, the specific process of step S5 includes: Sample input vectors are constructed based on meteorological characteristics, multiple time scale characteristics, and reconstructed sequences. , to construct the predicted values ​​of the sample input vectors for the XGboost model; Define the objective function of the XGBoost model, which includes a fitting loss function and a model complexity regularization term. The fitting loss function is the squared error between the true value and the predicted value of the reconstructed sequence, and the regularization term is the single tree regularization term of XGBoost for the reconstructed sequence. The objective function is trained and iterated until it is minimized and converged, thus obtaining the final prediction result.

[0015] A multi-scale load power prediction device based on multi-scale feature extraction includes: The preprocessing unit is used to preprocess the multivariate workload by removing outliers and filling in missing values ​​to obtain preprocessed data. The quantification unit is used to quantify the correlation and coupling characteristics between meteorological features and multivariate loads using Spearman rank correlation coefficients, and to screen out strongly correlated meteorological features; among which the coupling characteristics are the correlation with multivariate loads and the coupling characteristics between multivariate loads themselves. The reconstruction unit is used to decompose the preprocessed data using the CEEMDAN algorithm and reconstruct the sequence based on the sample entropy value of the decomposition result to obtain the reconstructed sequence. The extraction unit is used to extract the multi-time-scale variation patterns of preprocessed data based on coupling features and employing the Prophet model, which serves as multi-time-scale features. The prediction unit is used to input strongly correlated meteorological features, multiple time scale features, and reconstructed sequences into the XGboost model for prediction and fitting, and to obtain the final prediction results.

[0016] Furthermore, the reconfiguration unit includes: The pre-decomposition module is used to add different Gaussian white noise sequences to the preprocessed data to construct the pre-decomposition sequence; The initial decomposition module is used to perform EMD decomposition on the preprocessed data to obtain the first IMF component. The secondary decomposition module is used to preprocess the data, subtract the first IMF component to calculate the residual, and then perform EMD decomposition on the residual to obtain the remaining IMF components. This process is repeated until the residual is a monotonic sequence and no meaningful IMF components can be decomposed.

[0017] Furthermore, the extraction unit is used to extract potential multi-timescale fluctuation features in load power using the Prophet model. Specifically, it includes: using the Prophet model to divide the time series into four parts: a trend component module, a periodic component module, a holiday effect component module, and a noise component module. Among them: the trend component module represents the annual variation of load and is calculated using a nonlinear logistic regression function; the periodic component module represents the periodic characteristics of power load at different time scales and is constructed using Fourier series; the holiday effect component module represents the impact of holidays or sudden events on the series fluctuations and is modeled using holidays and special events; the noise component module represents unpredicted fluctuations and is represented using a normal distribution function.

[0018] Furthermore, the prediction unit includes: The input / output construction module is used to construct sample input vectors for strongly correlated meteorological features, multiple time scale features, and reconstructed sequences, and to construct the predicted values ​​of the sample input vectors of the XGboost model. The objective function construction module is used to define the objective function of the XGBoost model, which includes a fitting loss function and a model complexity regularization term. The fitting loss function is the squared error between the true value and the predicted value of the reconstructed sequence, and the regularization term is the single tree regularization term of XGBoost for the reconstructed sequence. The training module is used to train the objective function, iterating until the objective function is minimized and converges, to obtain the final prediction result.

[0019] The beneficial effects of this application include: Facilitates feature extraction: By using Prophet to decompose the load into trends, multi-cycles, and holiday effects, the inherent time evolution pattern of the load can be explicitly extracted. Unlike feature engineering that relies entirely on data-driven approaches, it can effectively capture load drift phenomena caused by seasonal changes and holidays, reducing the difficulty of non-stationary time series modeling.

[0020] Improved prediction accuracy: Compared with traditional prediction methods, the multi-scale load power prediction method based on multi-scale feature extraction adopted in this invention has higher prediction accuracy, providing a foundation for flexible power system dispatch. Attached Figure Description

[0021] Figure 1 The flowchart shows a multivariate load power prediction method based on multi-scale feature extraction. Figure 2 The results of the Spearman correlation analysis are shown in the figure. Figure 3 The image shows the results of multi-timescale feature extraction of electrical load. Figure 4 This is a diagram showing the results of multi-timescale feature extraction of cooling load; Figure 5 This is a graph showing the results of multi-timescale feature extraction of heat load. Figure 6 This is a graph showing the predicted electrical load. Figure 7 This is a graph showing the predicted cooling load. Figure 8 This is a graph showing the predicted heat load. Detailed Implementation

[0022] The above is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Example 1 Multivariate load power prediction methods based on multi-scale feature extraction, such as Figure 1 As shown, it includes the following steps: S1, Data Preprocessing: Neighbor-to-date imputation, a missing value handling technique, uses data from the days immediately before and after the time of the missing value as a reference, and employs mathematical calculations to reasonably estimate the missing value. The calculation process for neighbor-to-date imputation is shown below: (1) In the formula: Indicates the first sky Missing load data at any given time; , , , Indicates the correspondence between different adjacent days The weight of each moment; , , , Indicates the first The two days before and after Load data at different times. Corresponding times for different adjacent days. Weight of engraving Automatic selection can be achieved through a linear regression model.

[0024] Load data preprocessing is a core prerequisite for research and engineering applications such as power forecasting and grid dispatching. Its quality directly determines the effectiveness and reliability of subsequent model training, data analysis, and decision-making. Multivariate load data is influenced by multiple factors, including the natural environment, equipment status, and acquisition systems, and suffers from problems such as noise interference, missing anomalies, dimensional redundancy, and spatiotemporal heterogeneity. Directly using raw data can lead to model fitting bias, a sharp drop in prediction accuracy, dispatching decision errors, and even grid security risks. Therefore, data preprocessing is necessary.

[0025] S2 uses Spearman rank correlation coefficient to quantify the correlation and coupling characteristics between meteorological features and multivariate loads, and screens out strongly correlated meteorological features; among which the coupling characteristics are the correlation with multivariate loads and the coupling characteristics between multivariate loads themselves; The Spearman correlation coefficient method is a nonparametric statistical method primarily used to measure the monotonic relationship between two variables. In source-load power prediction research, the Spearman correlation coefficient method is a key tool for feature selection and variable correlation analysis. It improves upon the limitations of the Pearson correlation coefficient in scenarios with nonnormal or outlier data. It uses the Spearman rank correlation coefficient to measure the monotonic correlation between features and output load, enabling initial feature selection. Unlike the Pearson correlation coefficient, this method is not constrained by the distribution pattern of variables and is more suitable for the nonlinear and nonnormal characteristics of photovoltaic and load time-series data. After converting the original observations to their corresponding ranking ranks, substituting them into the Pearson correlation coefficient formula simplifies to the commonly used form, yielding the Spearman correlation coefficient: (2) In the formula, It is the first The rank difference of each sample It is the total number of samples for a feature. Generally speaking, It is believed that there is a weak correlation between the variables, with almost no actual monotonic association; It is believed that there is a moderate to weak correlation between the variables; It is believed that there is a moderate to strong correlation between the variables; It is believed that there is a strong correlation between the variables.

[0026] S3. The preprocessed data is decomposed using the CEEMDAN algorithm, and the sequence is reconstructed based on the sample entropy value of the decomposition result to obtain the reconstructed sequence. The decomposition steps of CEEMDAN are as follows: 1) For the original load sequence Add to For different Gaussian white noise sequences, construct pre-decomposition sequences: (3) In the formula, is No. Pre-decomposition sequence with added noise. For the first The white noise sequence added next time. This refers to the signal-to-noise ratio.

[0027] 2) To Perform EMD decomposition to obtain the first IMF component: (4) In the formula, This indicates that an EMD decomposition is performed on the signal within the parentheses.

[0028] 3) Subtracting the original sequence Calculate the residuals, and repeat step 2) iteratively to obtain the remaining IMF components: (5) (6) In the formula, For the first The components obtained from the second decomposition. This indicates that the signal within the parentheses will be executed. Secondary EMD decomposition.

[0029] 4) Repeat step 3) until the residual is a monotonic sequence from which no meaningful IMF components can be decomposed. The original signal can be represented as: (7) In the formula, The total number of components derived from the CEEMDAN decomposition. This is the final monotonic residual.

[0030] CEEMDAN decomposition typically generates numerous IMF components, and these components often exhibit some similarity. Predicting each component separately would significantly increase model training costs and amplify the overall prediction error due to the randomness of high-frequency noise components. Therefore, sample entropy (SE) is introduced as a measure of component complexity. Sample entropy is calculated as follows: (8) In the formula, For sequence dimensions; The total number of samples for the IMF component; For sequence dimensions Specific category The probability of it appearing in the sample set; For sequence dimensions hour The probability of appearing in the sample set. and It can characterize the properties of different IMF components.

[0031] Based on the sample entropy calculation results, K-Means clustering is introduced to recombine the frequency components, resulting in three recombined components of load power: a high-frequency random component, a mid-frequency periodic component, and a low-frequency trend component. These components respectively reflect the fluctuation characteristics of load power at different time scales. The K-Means clustering calculation formula is as follows: (9) In the formula, , It is the number of cluster sets. It is the first Cluster centers of each cluster.

[0032] Adaptive Noise Complete Ensemble Empirical Mode Decomposition (CEEMDAN) is one of the most widely used and adaptable algorithms in the field of power forecasting. It is an optimized iteration based on the EMD and EEMD algorithms, and its core purpose is to solve the problems of mode mixing, significant residual noise, and low computational efficiency of traditional decomposition methods. Its core principle is to adaptively decompose a complex non-stationary load power sequence into a series of single-frequency, stationary intrinsic mode functions (IMFs) and a final residual term through an adaptive noise addition strategy and a successive decomposition-averaging mechanism. Different IMF components correspond to short-term fluctuations in the power sequence (such as power abrupt changes caused by cloud cover and instantaneous wind speed changes), medium-term cycles (such as intraday load peaks and valleys and changes in photovoltaic output during the day), and the residual term corresponds to long-term trends (such as overall power fluctuations caused by seasonal changes).

[0033] S4. Based on coupling features, the Prophet model is used to extract the variation patterns of preprocessed data across multiple time scales, which are then used as multi-time scale features. To fully consider the multi-timescale trend characteristics of electricity load, this paper uses Prophet to extract potential multi-timescale trend features from load power. The extracted seasonal and trend features are then fused with strongly correlated meteorological factors as the input feature set. The Prophet model is a time series forecasting model developed by Facebook.

[0034] The Prophet model divides time series data into four components: trend, periodicity, holiday effect, and noise, including additive and multiplicative models. This paper chooses the additive model, with the specific formula as follows: (10) In the formula: A trend function representing the non-periodic changes in simulated time series values; Indicates periodic changes; This indicates that the effects of the holiday occur over a day or several days at a potentially irregular schedule; This is the residual term.

[0035] (1) Trend term function To characterize the annual variation in load, a nonlinear logistic regression function is used for calculation, and its expression is: (11) In the formula: It is the maximum capacity, i.e., the upper limit of the load. Indicates the basic rate of change. Indicates the bias amount. The change in rate of change This is to properly handle the offset adjustment introduced by the line segment boundary.

[0036] (2) Electricity load exhibits a clear periodicity across different time scales. For example, electrical load typically peaks in summer or winter, heating load typically peaks in winter, and cooling load typically peaks in summer; intraday load variations are typically at their lowest point in the early morning. (Periodic component) However, it typically uses weekly, monthly, or yearly time scales and is modeled based on Fourier series: (12) In the formula: The period value of the selected time scale, , is the smoothing parameter; P is the total number of Fourier decompositions, and p is the cyclic symbol.

[0037] (3) In time series forecasting, holidays and certain unforeseen events can significantly impact the forecast. Using these factors as prior knowledge for the model can effectively improve its prediction accuracy. Since holidays and special events do not follow a cyclical pattern, a specific matrix containing the dates and details of these holidays can be provided to model the holiday function. : (13) In the formula: This represents the total number of holidays; For the first A holiday window; 1 represents the load offset caused by the holiday; 1 is the indicator function. The value is 1 during holidays; otherwise, it is 0.

[0038] (4) This refers to the residual term, which represents the fluctuations not predicted by the model. It follows a normal distribution.

[0039] S5 inputs strongly correlated meteorological features, multiple time scale features, and reconstructed sequences into the XGboost model for prediction and fitting, and obtains the final prediction results.

[0040] XGBoost (Extreme Gradient Boosting) is an optimized ensemble regression and classification algorithm based on Gradient Boosting Decision Tree (GBDT), representing an engineering improvement upon GBDT. The model fits the target residual by serially stacking multiple decision trees. Through innovations such as multi-level regularization, second-order Taylor loss optimization, block-parallel computation, and automatic handling of missing values, it overcomes the shortcomings of traditional GBDT, including overfitting, slow convergence, and low computational efficiency. It is widely used in time series forecasting, load regression, and energy forecasting. For multivariate load forecasting tasks, XGBoost can efficiently mine the nonlinear coupling relationship between the multi-timescale components of the Prophet decomposition and the load, exhibiting strong fitting ability and interpretability.

[0041] Through the above steps, a feature set of meteorological features and multiple time-scale features, as well as a reconstructed sequence input set, can be constructed. For the input of the dataset constructed above, the first... sample vectors The XGBoost prediction output is: (14) In the formula: It corresponds to the first One IMF component sample The predicted value, Indicates the first An independent regression decision tree, It is the function space formed by all decision trees.

[0042] The objective function of XGBoost consists of two parts: a fitting loss function and a model complexity regularization term. These are used to control prediction error and suppress overfitting, respectively. The loss function represents the training error, and the regularization term defines the complexity of the tree. The smaller the value of the regularization term, the lower the model complexity and the stronger the generalization ability. XGBoost Objective Function The convergence condition can be defined as: (15) In the formula: It is the sample loss function, commonly calculated using the squared error; For the z-th IMF component sample, Indicates the first The regularization penalty term for each tree constrains the complexity of the tree structure; It is the leaf node splitting threshold; the larger it is, the simpler the tree is, and the more branches are suppressed. This represents the total number of leaf nodes on a single tree. is the leaf weight regularization coefficient, which constrains the output weight of the leaves; To indicate the number of trees The predicted score of each leaf node.

[0043] Example 2 A multi-scale load power prediction device based on multi-scale feature extraction includes: The preprocessing unit is used to preprocess the multivariate workload by removing outliers and filling in missing values ​​to obtain preprocessed data. The quantification unit is used to quantify the correlation and coupling characteristics between meteorological features and multivariate loads using Spearman rank correlation coefficients, and to screen out strongly correlated meteorological features; among which the coupling characteristics are the correlation with multivariate loads and the coupling characteristics between multivariate loads themselves. The reconstruction unit is used to decompose the preprocessed data using the CEEMDAN algorithm and reconstruct the sequence based on the sample entropy value of the decomposition result to obtain the reconstructed sequence. The extraction unit is used to extract the multi-time-scale variation patterns of preprocessed data based on coupling features and employing the Prophet model, which serves as multi-time-scale features. The prediction unit is used to input strongly correlated meteorological features, multiple time scale features, and reconstructed sequences into the XGboost model for prediction and fitting, and to obtain the final prediction results.

[0044] The reconstruction unit includes: The pre-decomposition module is used to add different Gaussian white noise sequences to the preprocessed data to construct the pre-decomposition sequence; The initial decomposition module is used to perform EMD decomposition on the preprocessed data to obtain the first IMF component. The secondary decomposition module is used to preprocess the data, subtract the first IMF component to calculate the residual, and then perform EMD decomposition on the residual to obtain the remaining IMF components. This process is repeated until the residual is a monotonic sequence and no meaningful IMF components can be decomposed.

[0045] The extraction unit is used to extract potential multi-timescale fluctuation features in load power using the Prophet model. Specifically, it includes: using the Prophet model to divide the time series into four parts: a trend component module, a periodic component module, a holiday effect component module, and a noise component module. Among them: the trend component module represents the annual variation of load and is calculated using a nonlinear logistic regression function; the periodic component module represents the periodic characteristics of power load at different time scales and is constructed using Fourier series; the holiday effect component module represents the impact of holidays or sudden events on the series fluctuations and is modeled using holidays and special events; the noise component module represents unpredicted fluctuations and is represented using a normal distribution function.

[0046] The prediction unit includes: The input / output construction module is used to construct sample input vectors for strongly correlated meteorological features, multiple time scale features, and reconstructed sequences, and to construct the predicted values ​​of the sample input vectors of the XGboost model. The objective function construction module is used to define the objective function of the XGBoost model, which includes a fitting loss function and a model complexity regularization term. The fitting loss function is the squared error between the true value and the predicted value of the reconstructed sequence, and the regularization term is the single tree regularization term of XGBoost for the reconstructed sequence. The training module is used to train the objective function, iterating until the objective function is minimized and converges, to obtain the final prediction result.

[0047] Application examples Figure 2 As shown in the Spearman correlation coefficient heatmap, heat load exhibits significant monotonic correlations with multiple meteorological variables. Heat load shows a strong negative correlation with temperature (-0.84) and precipitable water (-0.71), a moderate correlation with dew point temperature (-0.50) and air pressure (0.59), and meets the correlation threshold with relative humidity (0.46), solar zenith angle (0.33), and surface reflectance (-0.31). The absolute values ​​of the correlation coefficients for cloud cover, wind direction, and wind speed are all less than 0.3, indicating weak monotonic coupling.

[0048] The correlation characteristics of cooling load are significantly complementary to those of heating load. Cooling load shows a strong positive correlation with temperature (0.79) and precipitable water (0.68), and a moderate correlation with air pressure (-0.49), dew point temperature (0.47), and relative humidity (-0.41). Solar zenith angle (-0.31) and surface reflectance (0.38) also meet the screening criteria. Cloud cover, wind direction, and wind speed have correlation coefficients of less than 0.3, and their impact on cooling load is limited.

[0049] The correlation between electrical load and meteorological factors is generally weaker than that between heating and cooling loads. Electrical load shows a moderately strong positive correlation with temperature (0.62) and precipitable water (0.62), and a moderate correlation with dew point temperature (0.52) and air pressure (-0.47). Surface reflectance (0.31) just meets the screening threshold. The absolute values ​​of the correlation coefficients for relative humidity, solar zenith angle, cloud cover, wind direction, and wind speed are all below 0.3, indicating weak monotonic correlation.

[0050] Furthermore, analysis of the relationships between the comprehensive loads reveals a significant strong coupling characteristic among the electrical, cooling, and heating loads within the integrated energy system. The correlation coefficient between cooling and heating loads reaches -0.91, indicating a very strong negative correlation; the correlation coefficient between electrical and cooling loads is 0.81, showing a strong positive correlation; and the correlation coefficient between electrical and heating loads is -0.73, showing a relatively strong negative correlation. From a physical mechanism perspective, temperature changes are the core meteorological factor driving load evolution: under high-temperature conditions, cooling demand increases, leading to a simultaneous increase in cooling and electrical loads, while heating demand shrinks, resulting in a decrease in heating load; under low-temperature conditions, heating demand increases, causing heating load to rise while cooling and electrical loads correspondingly decline, ultimately forming a mutually reinforcing relationship among the three types of loads.

[0051] Figure 3 , Figure 4 , Figure 5 As shown, Figure 3 The purple curve represents the electrical load. Figure 4 The green curve represents the cooling load. Figure 5The red curve represents heat load. The results show that the three load types exhibit significant commonalities and differences across time scales. In the long-term trend dimension, electricity load shows a slow downward trend overall, cooling load initially rises and then falls, while heat load shows a slight but continuous increase, reflecting that the baseline evolution of the annual baseline load for the three load types is different. In the annual cycle dimension, cooling load exhibits a strong summer single-peak seasonal characteristic, while heat load shows a pattern of high in winter and low in summer, with both showing significant seasonal driving effects. Electricity load has a smaller annual cycle fluctuation amplitude and is less affected by seasonal factors than cooling and heat loads. In the weekly cycle dimension, all three load types show a pattern of high load on weekdays and lower load on weekends, with electricity and cooling loads showing significantly larger weekly fluctuation amplitudes than heat loads. In the daily cycle dimension, all three exhibit an intraday energy consumption rhythm of rising during the day and falling at night, with electricity and cooling loads showing prominent peak-to-valley differences, while heat load shows relatively gentle intraday fluctuations.

[0052] Multi-scale time-series decomposition results of electricity, cooling, and heating loads show that the coupling characteristics of the three types of loads have significant time-scale dependence. At annual, weekly, and intraday scales, cooling load and electricity load change phases are highly synchronized, exhibiting a strong positive coupling relationship. Peak electricity and cooling loads tend to overlap during midday in summer. Heating load exhibits negative coupling characteristics with electricity and cooling loads. In the annual cycle, heating load reaches its peak in winter while electricity and cooling loads are at a low level. Intraday, the peak heating load is concentrated in the early morning, offsetting the midday electricity and cooling load peaks, demonstrating a negatively correlated coupling effect. Heating load shows almost no significant fluctuations in the weekly cycle, but its trend is consistent with that of electricity and cooling loads on the weekly scale. In the long-term trend dimension, the evolution patterns of the three loads are different, and their driving factors are independent. Therefore, in comprehensive energy load forecasting modeling, it is necessary to fully utilize the positive coupling relationship between electricity and cooling loads, while also considering the negative coupling effect between heating load and electricity and cooling loads, and to characterize the interaction relationship of the three types of loads differently for different time scales.

[0053] Figure 6 , Figure 7 , Figure 8As shown, the results of independent single-load forecasting and integrated load forecasting considering the coupling relationship between electricity, cooling, and heat are compared. The time-series curves reveal that while single-load forecasting can roughly reproduce the overall load change trend, it deviates significantly at extreme points such as peaks and troughs. In contrast, the integrated load forecasting curve shows a higher degree of fit with the original data and better tracking of load peaks and troughs. Single-load forecasting only utilizes the historical information of the target load itself for modeling, ignoring the inherent coupling relationship between electricity, cooling, and heat loads in a multi-energy system, driven by meteorological conditions and user energy consumption behavior. Integrated load forecasting inputs multiple types of loads into the model, fully exploring the correlation between different energy flows and using the changing characteristics of other loads to assist in forecasting, effectively correcting prediction errors at extreme points. The accuracy gain from coupling information is more significant for electricity and heat loads, while the cooling load, with its strong temporal regularity, has a high baseline accuracy but a relatively limited improvement. These results demonstrate that introducing multi-load coupling characteristics in multi-energy load forecasting can effectively improve forecasting performance and better reflect the actual operating characteristics of multi-energy coupling in multi-energy systems.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes or alterations made by those skilled in the art using the disclosed technical content shall fall within the protection scope of the present invention.

Claims

1. A multivariate load power prediction method based on multi-scale feature extraction, characterized in that, include: S1, perform data preprocessing on the multivariate workload by removing outliers and filling in missing values ​​to obtain preprocessed data; S2, using Spearman's rank correlation coefficient to quantify the correlation and coupling characteristics between meteorological features and multivariate loads, and screening out strongly correlated meteorological features; The coupling characteristics include the correlation with multi-element loads and the coupling characteristics between multi-element loads themselves. S3. The preprocessed data is decomposed using the CEEMDAN algorithm, and the sequence is reconstructed based on the sample entropy value of the decomposition result to obtain the reconstructed sequence. S4. Based on coupling features, the Prophet model is used to extract the variation patterns of preprocessed data across multiple time scales, which are then used as multi-time scale features. S5 inputs strongly correlated meteorological features, multiple time scale features, and reconstructed sequences into the XGboost model for prediction and fitting, and obtains the final prediction results.

2. The multi-scale load power prediction method based on multi-scale feature extraction according to claim 1, characterized in that, The steps of the CEEMDAN algorithm are as follows: Different Gaussian white noise sequences are added to the preprocessed data to construct a pre-decomposition sequence; Perform EMD decomposition on the preprocessed data to obtain the first IMF component; The residual is calculated by subtracting the first IMF component from the preprocessed data. The residual is then decomposed into EMD to obtain the remaining IMF components. This process is repeated until the residual is a monotonic sequence from which no meaningful IMF components can be decomposed.

3. The multi-scale load power prediction method based on multi-scale feature extraction according to claim 2, characterized in that, The Prophet model was used to extract potential multi-timescale fluctuation characteristics in load power. Specifically, the time series was divided into four parts using the Prophet model: trend component, periodic component, holiday effect component, and noise component. The trend component represents the annual variation of load and is calculated using a nonlinear logistic regression function. The periodic component represents the periodic characteristics of power load at different time scales and is constructed using Fourier series. The holiday effect component represents the impact of holidays or sudden events on the series fluctuations and is modeled using a combination of holidays and special events. The noise component represents unpredictable fluctuations and is represented using a normal distribution function.

4. The multi-scale load power prediction method based on multi-scale feature extraction according to claim 2, characterized in that, The specific process of step S5 includes: Sample input vectors are constructed based on meteorological characteristics, multiple time scale characteristics, and reconstructed sequences. , to construct the predicted values ​​of the sample input vectors for the XGboost model; Define the objective function of the XGBoost model, which includes a fitting loss function and a model complexity regularization term. The fitting loss function is the squared error between the true value and the predicted value of the reconstructed sequence, and the regularization term is the single tree regularization term of XGBoost for the reconstructed sequence. The objective function is trained and iterated until it is minimized and converged, thus obtaining the final prediction result.

5. A multi-scale load power prediction device based on multi-scale feature extraction, characterized in that, include: The preprocessing unit is used to preprocess the multivariate workload by removing outliers and filling in missing values ​​to obtain preprocessed data. The quantification unit is used to quantify the correlation and coupling characteristics between meteorological features and multivariate loads using Spearman's rank correlation coefficient, and to screen out strongly correlated meteorological features. The coupling characteristics include the correlation with multi-element loads and the coupling characteristics between multi-element loads themselves. The reconstruction unit is used to decompose the preprocessed data using the CEEMDAN algorithm and reconstruct the sequence based on the sample entropy value of the decomposition result to obtain the reconstructed sequence. The extraction unit is used to extract the multi-time-scale variation patterns of preprocessed data based on coupling features and employing the Prophet model, which serves as multi-time-scale features. The prediction unit is used to input strongly correlated meteorological features, multiple time scale features, and reconstructed sequences into the XGboost model for prediction and fitting, and to obtain the final prediction results.

6. The multi-scale load power prediction device based on multi-scale feature extraction according to claim 5, characterized in that, The reconstruction unit includes: The pre-decomposition module is used to add different Gaussian white noise sequences to the preprocessed data to construct the pre-decomposition sequence; The initial decomposition module is used to perform EMD decomposition on the preprocessed data to obtain the first IMF component. The secondary decomposition module is used to preprocess the data, subtract the first IMF component to calculate the residual, and then perform EMD decomposition on the residual to obtain the remaining IMF components. This process is repeated until the residual is a monotonic sequence and no meaningful IMF components can be decomposed.

7. The multi-scale load power prediction device based on multi-scale feature extraction according to claim 5, characterized in that, The extraction unit is used to extract potential multi-timescale fluctuation features in load power using the Prophet model. Specifically, it includes: using the Prophet model to divide the time series into four parts: a trend component module, a periodic component module, a holiday effect component module, and a noise component module. Among them: the trend component module represents the annual variation of load and is calculated using a nonlinear logistic regression function; the periodic component module represents the periodic characteristics of power load at different time scales and is constructed using Fourier series; the holiday effect component module represents the impact of holidays or sudden events on the series fluctuations and is modeled using holidays and special events; the noise component module represents unpredicted fluctuations and is represented using a normal distribution function.

8. The multi-scale load power prediction device based on multi-scale feature extraction according to claim 5, characterized in that, The prediction unit includes: The input / output construction module is used to construct sample input vectors for strongly correlated meteorological features, multiple time scale features, and reconstructed sequences, and to construct the predicted values ​​of the sample input vectors of the XGboost model. The objective function construction module is used to define the objective function of the XGBoost model, which includes a fitting loss function and a model complexity regularization term. The fitting loss function is the squared error between the true value and the predicted value of the reconstructed sequence, and the regularization term is the single tree regularization term of XGBoost for the reconstructed sequence. The training module is used to train the objective function, iterating until the objective function is minimized and converges, to obtain the final prediction result.

Images