Dynamic hierarchical granularity model power load prediction method considering time sequence factors

Abstract

The invention relates to power load prediction, in particular to a dynamic hierarchical granularity model power load prediction method considering time sequence factors. According to the method, the fitting capability of periodic features, trend changes and sudden influences in load fluctuation is remarkably enhanced, and the problem of feature loss caused by neglect of a time sequence dynamic weight of an existing model can be solved. Comprising the following steps: S1, performing time sequence characteristic analysis on historical power load data, and converting an original non-stationary time sequence into a stationary time sequence through difference and logarithm transformation; s2, obtaining a cleaned power load data set; s3, based on the data cleaned in S2, selecting a Gaussian radial basis function (RBF) as a kernel function, and analyzing temperature, holiday identification and linear change trend key influence factors at the same time; s4, performing dynamic multi-level granulation on the data set, and dynamically adjusting the particle level according to the data mixing degree and the particle density; and S5, fusing the time sequence kernel function and the influence factor through a decision function, and predicting the future power load.

Description

Technical Field

[0001] This invention relates to power load forecasting, and more particularly to a dynamic hierarchical granular model power load forecasting method that considers time-series factors. Background Technology

[0002] As a core supporting technology for the safe dispatch of power systems, the accuracy of power load forecasting directly affects the economic operation and risk prevention of power grids with a scale of hundreds of millions of units. However, under the current trend of high penetration of new energy sources and diversified loads, power load data exhibits strong nonlinearity, non-stationarity, and multi-scale time-series coupling characteristics, posing a severe challenge to traditional forecasting methods when dealing with such dynamic systems.

[0003] The main limitations of traditional methods are:

[0004] Backpropagation Neural Network (BPNN): Although it has nonlinear mapping capabilities, its hyperparameter sensitivity and convergence instability make it difficult to adapt to time-varying load distributions.

[0005] Optimization models such as the Artificial Bee Colony Algorithm (ABC) are too complex to meet the minute-level response requirements for short-term predictions due to excessive parameter tuning complexity.

[0006] Long Short-Term Memory (LSTM) networks: While they can capture temporal dependencies, they are extremely sensitive to random noise, and the large number of parameters creates a huge computational bottleneck.

[0007] These shortcomings together result in high fluctuations in prediction errors of existing methods in complex scenarios, which seriously restricts the refined management and control of the power grid.

[0008] In Support Vector Machine (SVM) methods, the core challenge lies in the lack of a dynamic temporal mechanism. While traditional granular SVM (GSVM) improves processing efficiency through data granulation, its static hierarchical strategy (such as fixing subsets by amplitude) completely ignores the evolution of feature values ​​and their weight changes over time. For example:

[0009] The weight of air conditioning load in the summer afternoon needs to be dynamically increased, but GSVM still treats it as low-value data, resulting in the dilution of the contribution of key time series patterns (such as sudden changes in the morning peak).

[0010] While some improved solutions introduce a feature extraction module, they fail to establish a weight-temporal linkage mechanism, leading to the following problems:

[0011] The fixed feature weights prevent the differentiated weighting of features at different time scales (minute fluctuations / weekly cycles).

[0012] A rigid hierarchical structure prevents sub-granularity sets from migrating across levels as weights change.

[0013] This architecture defect makes low-value noise data (such as early morning load fluctuation) continue to occupy computing resources, while high-value features (such as the steep rise of load when the cold wave comes) cannot be dynamically enhanced, and finally the average identification accuracy of the model at the load turning point is less than 65%. SUMMARY

[0014] The present application is aimed at the defects in the prior art, and provides a dynamic hierarchical granularity model power load prediction method considering time sequence factors.

[0015] To achieve the above-mentioned purpose, the present application adopts the following technical scheme, a dynamic hierarchical granularity model power load prediction method considering time sequence factors, comprising the following steps:

[0016] S1, time sequence feature analysis is performed on historical power load data, the power load data is constructed into an equal-interval time sequence data set, the original non-stationary time sequence is converted into a stationary time sequence through difference and logarithmic transformation, an autoregressive model (AR model) is constructed based on the stationary sequence, and the time dependence and periodicity of the power load are extracted;

[0017] S2, the historical power load data is preprocessed and abnormal value cleaned, the abnormal data is identified as an outlier, and the outlier is repaired based on a time proximity weighted average method to obtain a cleaned power load data set;

[0018] S3, based on the data cleaned in S2, a Gaussian radial basis function (RBF) is selected as a kernel function, a time sequence kernel function is constructed in combination with the time sequence characteristics (time correlation) of the power load time sequence; meanwhile, temperature, holiday identification and linear trend key influence factors are analyzed, and the correlation between them and the load change is quantified;

[0019] S4, based on the time sequence kernel function and the influence factors, the data set is dynamically multi-layered, and the particle level is dynamically adjusted according to the data mixing degree and the particle density;

[0020] S5, based on the data set after the particle layering in S4, a time sequence sensitive load prediction model is constructed, the time sequence kernel function and the influence factors are fused through a decision function, and the future power load is predicted.

[0021] The step S1 specifically comprises:

[0022] S11, the power load data set is sampled and recorded according to equal-length time intervals to form an equal-interval time sequence data set;

[0023] S12, the equal-interval time sequence data set obtained in S11 is subjected to stationarity test, when it is determined to be a non-stationary time sequence, it is converted into a stationary time sequence through difference and logarithmic transformation;

[0024] S13, based on the stationary time series converted in S12, an autoregressive model (AR) is constructed, and the load value X t of the previous m time intervals is set as the historical load value X t-m t-m+1 t-1 , and the model expression is:

[0025]

[0026] wherein, is the autoregressive coefficient, and α t is the residual sequence, and X t is the load value at the current time;

[0027] S14: The order m of the model is determined by the autocorrelation function (ACF) analysis, and the strongest time lag of autocorrelation is selected as the main influence order; wherein, m = 24; the least square method is used to estimate the autoregressive coefficients of each order, and the statistical characteristics are analyzed to reveal the time sequence dependence structure of the load.

[0028] S2 specifically includes:

[0029] S21, the historical power load data set is divided into several sub-data sets according to the same year and the same month, and the data with a load value of 0 is set as an outlier;

[0030] S22, the average value of the hourly load in each sub-data set is calculated, and the data less than 1 / 4 of the average value is set as an outlier;

[0031] S23, the weighted average of all data in the sub-data set except the outliers is calculated, and if the absolute percentage error (APE) of the hourly load value and the weighted average is greater than 50%, the hourly load value is marked as an outlier;

[0032] S24, for the outliers in each sub-data set, the weighted average of the 7 days before and after the outliers is used to replace the load data of the outliers, and the calculation formula is:

[0033]

[0034] wherein, is the weighted average of the load value of the outlier t, is the load prediction value of the 7 days before and after the discrete point, i.e. the range of 14 days before and after the discrete point, t is the time point set as the outlier, and f t-i , f t+i is the weight value, and the weight value ranges from 7 to 1 according to the distance from the outlier time, and is calculated by f t-i = f t+i = (8-i);

[0035] ​​S25, after all outliers are repaired, integrate each sub-data set to form a complete and cleaned power load time series data set.

[0036] S3 specifically comprises:

[0037] S31, select Gaussian radial basis function (RBF) as the basic kernel function, and its expression is:

[0038]

[0039] wherein, x i , x j is the sample point in the power load time series, and σ is the kernel width parameter, which is used to adjust the sensitivity of the Gaussian kernel function to the distance between samples;

[0040] S32, based on the RBF kernel function, construct a time series kernel function combined with the time series dependence of the power load, specifically comprising:

[0041] Divide the feature space into m windows T={W1, W2,..., W m ,} with time points as the center, and each window corresponds to a different time neighborhood.

[0042] The window correction function and the correction kernel function are respectively shown in the following formulas:

[0043]

[0044] Define the time series kernel function as:

[0045] S33, perform influence factor analysis and quantification;

[0046] S3301, according to the time series characteristics of the power load, determine the following key influence factors: current hour temperature T t , average temperature of the previous and next 3 hours T t-3 , average temperature of the past 24 hours T a , monthly maximum temperature T max , holiday identifier Hs, and temperature linear change trend Trend.

[0047] S3302, construct a factor function Factor(θ) to quantify the contribution of each factor to the power load, and the calculation formula is:

[0048] Factor(θ)=θ1Trend+θ2(η1T max +η2T t +η3T t-3 +η4T a )+θ3Hs

[0049] s.t.θ1+θ2+θ3=1,η1+η2+η3+η4=1

[0050] Wherein, the typical value of trend value coefficient θ1 is 0.75, the typical value of temperature coefficient θ2 is 0.15, the temperature sub-coefficients η1=0.05, η2=0.7, η3=0.15, η4=0.1, and the typical value of holiday coefficient θ3 is 0.1.

[0051] The step 4 specifically comprises:

[0052] S41, based on the time window weight of the time sequence kernel function and the feature importance score of the influence factor, the data set is divided into information particles with different mixed degrees;

[0053] S42, according to the time sequence correlation strength of the information particle, that is, the sensitivity degree of the information particle, the kernel width parameter σ in the time sequence kernel function and the contribution value of the influence factor are determined, and real-time migration to different granular levels is performed;

[0054] S43, the information particles with high time sequence sensitivity are divided in fine granularity, and the information particles with low sensitivity are divided in coarse granularity.

[0055] The step S5 comprises:

[0056] S51, based on the dynamic multi-level granulation data set obtained in S4, the training set and the test set are divided in time sequence, and the data is normalized;

[0057] S52, based on the data set processed in S51, a time sequence sensitive load prediction model is constructed, wherein the time sequence kernel function defined in S3 is used as the model kernel function, and the influence factors (temperature, holiday identifier and trend factor) quantified in S3 are used as the model input features; the decision function is in the form of:

[0058]

[0059] Wherein, is a support vector coefficient, and b * is a bias term;

[0060] S53, the test set is input into the trained model, high weight is assigned to the data of high granular level (fine granularity), low weight is assigned to the data of low granular level (coarse granularity), the future power load is predicted by using the decision function, and the output result is a time sequence continuous load value sequence.

[0061] Compared with the prior art, the present application has the following beneficial effects.

[0062] The application can significantly improve the accuracy and reliability of power load prediction. By constructing a dynamic hierarchical granularity framework, the load data is granulated into multiple sub-granularity sets according to characteristic values and distributed to different weight levels, and the weight proportion and level distribution of each sub-granularity are dynamically adjusted combined with time sequence factors, effectively capturing the nonlinear, non-stationary characteristics and time sequence evolution law of power load data. Compared with traditional methods, the technical scheme of the application performs better in core indicators such as MASE (mean absolute standardized error) and MAPE (mean absolute percentage error), especially the fitting ability of periodic characteristics, trend changes and sudden impacts in load fluctuations is significantly enhanced, solving the feature loss problem caused by ignoring time sequence dynamic weight in existing models. BRIEF DESCRIPTION OF DRAWINGS

[0063] The application will be further described below in combination with the drawings and specific embodiments. The protection scope of the application is not limited to the following content.

[0064] Figure 1 is a schematic diagram of statistical characteristics of original training set data.

[0065] Figure 2 is a schematic diagram of statistical characteristics of load data set after data cleaning and outlier processing.

[0066] Figure 3 is a graph of initial granulation number, initial granulation level number and dynamic granulation parameter change.

[0067] Figure 4 is a flowchart of granulation and dynamic hierarchical adjustment algorithm.

[0068] Figure 5 is a schematic diagram of a dynamic hierarchical granularity power load prediction model considering time sequence factors.

[0069] Figure 6 is a graphical comparison of power load prediction results. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the present disclosure will be described clearly and completely in combination with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present disclosure.

[0071] The terms used in the embodiments of the present disclosure are only for the purpose of describing specific embodiments, and are not intended to limit the present disclosure. The singular forms "a", "an" and "the" used in the embodiments of the present disclosure and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0072] Depending on the context, words such as “if” or “suppose” used here can be interpreted as “when”, “when”, “in response to determination”, or “in response to detection”.

[0073] For ease of understanding, the embodiments of this disclosure will be described in detail first.

[0074] like Figures 1-6 As shown, step S1: Perform time-series characteristic analysis on historical power load data.

[0075] S11: The power load dataset is sampled and recorded at equal time intervals to form a typical equal-interval time series dataset. In a specific implementation, the dataset used is the GEFCom2012 dataset provided by the IEEE Energy Forecasting Working Group (WGEF). In a specific implementation example, power load data from 2004 to 2007 was selected and sampled and recorded at equal time intervals of 1 hour to form an equal-interval time series dataset. This time series exhibits typical non-stationary characteristics: statistically, the mean of the power load data from 2004 to 2007 is 23103, 18369, 18165, and 19966, respectively, and the standard deviations are 5665, 5860, 5323, and 6232, respectively. Both the mean and variance change with the year. At the same time, the change of its autocorrelation function depends on the time interval k. For example, the correlation between adjacent hours is significantly different from the correlation between 24-hour intervals, which conforms to the dynamic characteristics of non-stationary time series.

[0076] S12: The time series of GEFCom2012 power load data is transformed into a stationary time series model using differencing and logarithmic transformation methods. To transform the above non-stationary time series into a stationary series suitable for autoregressive (AR) models, the hourly power load data from 2004 to 2007 is processed as follows:

[0077] Take the natural logarithm of the original load data to reduce the difference in standard deviation between different years, such as Std.Dev = 5665 in 2004 and Std.Dev = 6232 in 2007, which would lead to variance nonstationarity and stabilize the fluctuation range of the series.

[0078] The first difference of the logarithmically transformed sequence is performed to eliminate the trend in the data, such as the annual trend of the mean fluctuating from 23103 to 19966 from 2004 to 2007, so that the mean of the sequence tends to be stable. After the above processing, the power load time series is transformed into a stationary series, which can be used to construct an autoregressive (AR) model.

[0079] S13: Based on the stationary time series processed in S12, an autoregressive (AR) model is constructed using hourly electricity load data from 2004 to 2007 in the GEFCom2012 dataset. Let the current system observation value X be... t That is, the historical observations X for the preceding m time intervals. t-m ,X t-m+1 ,...,X t-1 A linear combination of these factors allows the AR model to be used with the following formula: Obtained from the formula α is the autoregressive coefficient. t It is a residual sequence.

[0080] Based on the statistical characteristics of the experimental data, analysis using the autocorrelation function (ACF) revealed that the autocorrelation of power load was most significant at the 24-hour interval (daily cycle) (correlation coefficient 0.71), consistent with the periodic characteristics of daily load data (e.g., the daily average load in 2004 was 427,065, with a maximum of 838,313; the daily average load in 2007 was 479,174, with a maximum of 820,200, both showing obvious intraday fluctuations). Therefore, the model order was determined to be m = 24, focusing on analyzing the impact of historical load values ​​from the previous 24 hours on the prediction results.

[0081] Using stationary series data from 2004 to 2007, the autoregressive coefficients were estimated using the least squares method. to The coefficient corresponding to time t-24 (the same hour of the previous day) The absolute value (0.63) is the largest, consistent with the stability characteristics of hourly data (e.g., the hourly average fluctuation range from 18165 to 23103 and the median from 16989 to 23219 from 2004 to 2007); the coefficient of adjacent time periods (e.g.) The small absolute value (0.15) reflects the weak load correlation within a short time interval, which is consistent with the random fluctuation characteristics reflected by the hourly data standard deviation (5323-6232), and reflects the time series characteristics of the power load dataset.

[0082] Table 1 Statistical characteristics of the load dataset

[0083]

[0084] Step S2: Perform data preprocessing and outlier cleaning on historical power load data;

[0085] S21: Divide the dataset to be analyzed into several subsets based on the same year and the same month, and set the data with a loading value of 0 as outliers;

[0086] S22: Calculate the average value of each hour load in the sub-data set, and set all data less than 1 / 4 of the average value as outliers;

[0087] S23: Calculate the weighted average of all data excluding outliers, and if the absolute percentage error (APE) of the hourly load value of a certain data from the weighted average is greater than 50%, it is set as an outlier;

[0088] S24: After identifying all outliers in each sub-data set, replace the load data of the outliers with the weighted average of the 7 days before and after the outliers, as shown in the following formula:

[0089]

[0090] S25: In the formula shown in step S24, is the weighted average of the outlier t load value, is the load prediction value of the 7 days before and after the outlier, i.e. the range of 14 days before and after the outlier, t is the time point set as the outlier, f t-i , f t+i is the weight value, which ranges from 7 to 1 according to the distance from the outlier time, and is calculated as f t-i = f t+i = (8-i) The load data set after data preprocessing and outlier cleaning is shown in the following figure.

[0091] Further, when performing step S3, first select the time series kernel function, which has the following specific form:

[0092] S31: Based on the time series characteristics of power load prediction, select Gaussian radial basis function as the kernel function, which can also be called (Radial Basic Function, RBF) kernel function. The RBF kernel function has the characteristics of fast global convergence, which is shown in the following formula:

[0093]

[0094] S32: Based on RBF and function, construct a time series kernel function, the steps are as follows:

[0095] [1] Based on the time correlation characteristics of time series data, establish a time kernel function. The feature space formed by the time series data is divided into several windows with the time point as the center. The window covers the complete feature space and different windows can overlap each other. The granular points with similar time attribute values can appear in the same window, otherwise they cannot appear in the same window.

[0096] [2] Use the training set samples with the time point as the center, divide the feature space into m windows T = {W1, W2,..., Wm ,},window correction function and correction kernel function are respectively shown as follows:

[0097] [3]

[0098] [4] define the timing kernel function as

[0099] Further, when step S4 is performed, the power load data set containing the timing feature values is granulated using a granulation algorithm, and the granulation level at which the particles are located is adjusted according to a particle dynamic adjustment strategy, to obtain a sample data set for power load prediction;

[0100] S41: the power load data set containing the timing feature values is granulated using a multi-level granulation algorithm, and the specific steps are as follows:

[0101] [1] the initial data set is subjected to a primary granulation process to obtain a granulated initial granulation set H;

[0102] [2] the mixing degree, density, center and radius of each particle in the initial granulation set are calculated;

[0103] [3] the center of each particle in H is added to the training set for training, to obtain a classification hyperplane f(x) = sgn(W*φ(x) + b), and the classification accuracy of the particle is calculated, if the accuracy condition is met, the granulation is stopped, and [6] is entered, otherwise [4] is entered;

[0104] [4] the dynamic granulation factor of each particle is calculated and the mixing degree I(H k ), according to the values of and I(H k ), the particles that need to be divided again are found to enter the next level of granulation process, if all particles do not meet the condition to enter the next level of granulation, [6] is entered;

[0105] [5] the divided particles replace the original particles in the initial granulation set to generate a new granulation set, and [6] is entered;

[0106] [6] the optimized hyperplane model f'(x) = sgn(W*φ(x) + b) of the particle set at different levels is output, and the algorithm ends.

[0107] S42: the granulation level at which all particles in the particle set are located is dynamically adjusted using a multi-level clustering division strategy, and the specific steps are as follows:

[0108] [1] Set multiple hyperplanes, each corresponding to a different weight range, and convert the hierarchical space division into multiple weight range hyperplane division; the particle determines its belonging to a specific hyperplane, i.e., a specific granular level, according to the weight value carried by the particle, i.e., the particle density ρ i value of the particle. i The calculation method of the particle density ρ i value is as follows:

[0109] 1) Given a training set sample The positive and negative class samples in the training set are divided into K information granules by the K-means clustering algorithm, and the center and radius of a certain granule G i and r i ,

[0110] 2) μ i and r i are expressed by the following formulas:

[0111]

[0112] 3) The distance of the particle sample to any granule G i center μ i can be expressed as:

[0113] 4)

[0114] 5) Let the hyperplane f of the granular support vector machine be: G i is a granule obtained by using the granular division method, and the center point μ i is selected as the representative point to calculate the distance D between the granule G i and the hyperplane. When the granule G i is near the hyperplane, it is a strong information granule; otherwise, it is a weak information granule.

[0115] 6) As shown in the following formula,

[0116]

[0117] 7) The particle density is expressed as follows: assuming that there is a granule G i ={x i}(i=1,2,...,l), containing n i samples, and its radius and center are r i and μ i , respectively. The density ρ i of the granule G i is defined as follows:

[0118]

[0119] [2] Update the particle weight value over time, and dynamically migrate the particle from the original hyperplane to the new hyperplane based on the updated value to realize dynamic adjustment of the level.

[0120] [3] Perform strong and weak division and level adjustment on information particles: strong information particles (large sample size in the region and vicinity, and large influence on classification) are trained in finer levels to improve sample set training accuracy; weak information particles (small sample size in the region and vicinity) reduce the training level to improve training efficiency.

[0121] [4] By dynamically granulating different particle density data sets, different density granulation data sets are obtained, and information particles are divided into different levels in the multi-dimensional space according to their density, that is, the sample data set for power load prediction is obtained.

[0122] Further, when performing step S5, first determine the load prediction model parameters, the model decision function and the related conditions; generate the load prediction training set data; input the training set data into the load prediction model, and complete the power load prediction.

[0123] S51: Determine the load prediction model parameters, the decision function and the related conditions

[0124] [1] The granulation process and granulation level division of information particles are the most important instance links. The main parameters for the granulation process and granulation level division include the initial granulation number K, the initial granulation level number Lev, and the d_para dynamic granulation parameter. Different types of data sets may correspond to different optimal values of the above three parameters, and the values of the above three parameters need to be adjusted according to the data set used in the specific instance.

[0125] [2] In combination with the training set data, the initial granulation number K, the initial granulation level number Lev, and the d_para dynamic granulation parameter are analyzed using the granulation process and dynamic level adjustment method in step S4. The specific instance analysis results are shown in Tables 1, 2 and Figure 3

[0126] Table 1: Power load data parameters in January 2004 and January 2005

[0127]

[0128]

[0129] Table 2: Power load data parameters in January 2006 and January 2007

[0130]

[0131] ​[3] When the dynamic granularity parameter d_para = 1, Lev_over = 10, and the initial granulation number K = 10 are selected, the granulation number tends to be stable with the increase of Lev_over, that is, the prediction accuracy is the highest, as shown in the blue line in Fig. 3. Figure 3 Considering the influence of different parameter values on the prediction accuracy as described above, in order to ensure the accuracy of the prediction, the subsequent specific example process will use the parameter combination shown by the blue line in Fig. 3, that is, d_para = 1, Lev_over = 10, and k = 10. Figure 3 Considering the influence of different parameter values on the prediction accuracy as described above, in order to ensure the accuracy of the prediction, the subsequent specific example process will use the parameter combination shown by the blue line in Fig. 3, that is, d_para = 1, Lev_over = 10, and k = 10.

[0132] S52: Data granulation and dynamic hierarchy adjustment are performed on the initial training set using the initial granulation number K, the initial granulation hierarchy number Lev, and the d_para dynamic granularity division parameter value determined in step S51 to ensure the effective accuracy of the load prediction, and the specific steps are as follows:

[0133] [1] Initialize the sample set and initialize the parameters according to the preset values, including the granularity number, the granularity hierarchy, and the dynamic granularity division parameter;

[0134] [2] Granulate the sample set X0 in the initial training set T0. Set the current level and the granulation result.

[0135] [3] Update the current granulation level, extract the granular core in the granulation layer, and obtain the approximate regression hyperplane of the granulation layer using the training set.

[0136] [4] Calculate the distance of each particle in the Lev layer to the approximate regression plane f Lev of the layer, and update the information granules on the Lev granulation layer according to the calculated distance value.

[0137] [5] Calculate the density of the information granules on the updated Lev granularity layer, calculate the K value on the Lev granularity layer, which represents the dynamic granularity number of the jth information granule on the Lev granularity layer, and update the information granule data set on the Lev granularity layer.

[0138] [6] If the current level reaches the preset value, step 7 is performed. Otherwise, steps 3 to 5 are repeated until the current level reaches the preset value.

[0139] [7] Obtain the information granules at each level as the sample data set of the load prediction model, realize the power load data prediction, and obtain the final prediction result. The specific process is shown in Fig. 4: Figure 4

[0140] S53: Use the prediction training set input obtained in step S52 to execute the load prediction process and obtain the load prediction result:

[0141] ​[1] The load prediction model architecture mainly includes a data cleaning and preprocessing layer, a data feature analysis and extraction layer, a data granularity and dynamic adjustment layer, and a data prediction model training and verification layer, each layer contains specific functions to be completed, as shown in Figure 5

[0142] [2] To verify the effectiveness of the method, the power load data from January 1, 2008 to January 15, 2008 is used as the verification set data, and the load prediction results obtained by the classical granularity support vector machine (GSVM), kernel granularity support vector machine (KGSVM), back propagation neural network (BPNN), artificial bee colony (ABC), long short term memory (LSTM) method and the method are compared, and the comparison results are shown in Table 3 and Figure 6

[0143] Table 3 Numerical comparison of power load prediction results

[0144]

[0145] [3] To test the prediction performance of the method, the mean absolute percentage error (MAPE) and the mean absolute standardized error (MASE) are used as the prediction index. MAPE is used to compare the percentage error between the true value and the predicted value, and the smaller the MAPE, the closer the predicted value to the true value. MASE is a measure of prediction accuracy, and its value ranges from 0 to 1. The smaller the value of MASE, the higher the prediction accuracy. The comparison results are shown in Table 4, and it can be seen from Table 4 that the prediction index of the method is higher than that of the current common load prediction method.

[0146] Table 4 MASE and MAPE prediction index values of different models

[0147] Model name MASE MAPE% GSVM 0.759 3.225 KGSVM 0.704 3.032 BPNN 0.738 3.119 ABC 0.642 2.870 LSTM 0.703 2.751 DHGSVM-TF 0.634 2.649

[0148] The scheme has stronger robustness and generalization ability, and can adapt to complex and variable power load scenarios. Through the dynamic hierarchical adjustment mechanism, the algorithm can automatically optimize the granularity level according to the evolution of data features over time, and the sensitivity to noisy data and outliers is significantly reduced. By discarding low-weight particles to simplify the calculation complexity, the processing efficiency of large-scale data sets is improved while ensuring the prediction accuracy. Even in the case of insufficient data or local missing, the stable prediction performance can still be maintained, meeting the strict requirements of load prediction model reliability in actual engineering.

[0149] ​​The scheme effectively reduces the calculation cost and implementation difficulty of power load prediction. The dynamic hierarchical granulation process realizes the fine division of the feature space through weight differentiation, reduces the interference of redundant data on model training, and the parameter optimization mechanism is more concise and requires less hardware resources, which can be efficiently run on a regular computing platform. This feature makes it easier to deploy in power system dispatch centers, power distribution terminals and other practical scenarios, providing a prediction tool with high precision and low cost for grassroots power departments.

[0150] The scheme provides strong support for the fine scheduling and decision of the power system. The high-precision load prediction results can be directly applied to power production planning, unit start-stop optimization, power flow regulation, etc., which helps to reduce power generation cost, reduce wind and light curtailment, and improve the safety and economy of power grid operation. Compared with traditional methods, the prediction accuracy of short-term load fluctuations is greatly improved, which can better adapt to new energy grid connection, user-side response and other new power system scenarios, and provides technical support for load management under the background of power market reform.

[0151] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "illustrative embodiment", "preferred embodiment", "specific implementation" or "preferred implementation" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0152] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand; it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; thus, these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present application.

Claims

1. A dynamic hierarchical granular model method for power load forecasting that considers time-series factors, characterized in that, Includes the following steps: S1. Perform time series feature analysis on historical power load data, construct the power load data into an equally spaced time series dataset, transform the original non-stationary time series into a stationary time series through difference and logarithmic transformation, construct an autoregressive model based on the stationary series, and extract the time dependence and periodic characteristics of power load. S2. Perform preprocessing and outlier cleaning on historical power load data, identify outliers as outliers, and repair outliers based on time proximity weighted average method to obtain cleaned power load dataset. S3. Based on the data cleaned in S2, the Gaussian radial basis function is selected as the kernel function, and a time series kernel function is constructed by combining the time series characteristics of the power load time series; at the same time, the key influencing factors of temperature, holiday indicators and linear change trends are analyzed, and their correlation with load changes is quantified. S4. Based on the time series kernel function and the influence factor, the dataset is dynamically granulated into multiple levels, and the particle level is dynamically adjusted according to the data mixing degree and granularity. S5. Based on the granularized dataset from S4, a time-sensitive load forecasting model is constructed. By fusing the time-series kernel function and influencing factors through a decision function, the future power load is predicted.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: S11. Sample and record the power load dataset at equal time intervals to form an equal-interval time series dataset; S12: Perform a stationarity test on the equal-interval time series dataset obtained in S11. If it is determined to be a non-stationary time series, transform it into a stationary time series through difference and logarithmic transformation. S13. Based on the stationary time series transformed in S12, construct an autoregressive model, assuming the current load value X... t X represents the historical load values ​​for the previous m time intervals. t-m ,X t-m+1 ,...,X t-1 The linear combination of the two is expressed as follows: in, α is the autoregressive coefficient. t Let X be the residual sequence. t This represents the load value at the current moment. S14: The model order m is determined by autocorrelation function analysis, and the time lag with the strongest autocorrelation is selected as the main influence order; where m = 24; the autoregression coefficients of each order are estimated by the least squares method, and their statistical characteristics are analyzed to reveal the time-dependent structure of the load.

3. The method according to claim 1, characterized in that: S2 specifically includes: S21. Divide the historical power load dataset into several subsets based on the same year and the same month, and set the data with a load value of 0 as outliers. S22. Calculate the hourly average load in each subset of data, and set out the outliers for data that are less than 1 / 4 of the average load. S23. Calculate the weighted average of all data in the subset except for outliers. If the absolute percentage error between the hourly load value and the weighted average is greater than 50%, then mark it as an outlier. S24. For outliers in each subset, replace the outlier's load data with the weighted average of the 7 days before and after the outlier. The calculation formula is as follows: in, The weighted average of the load values ​​at outlier point t. The load forecast is for 7 days before and after the outlier, i.e., a total of 14 days before and after the outlier, where t is the time point set as the outlier, and f is the load forecast value for the period 7 days before and after the outlier. t-i f t+i The weights are determined by their temporal distance from outliers, ranging from 7 to 1, and are determined by f. t-i =f t+i = (8-i) is calculated to obtain; S25. After repairing all outliers, integrate the various subset datasets to form a complete and cleaned power load time series dataset.

4. The method according to claim 1, characterized in that: S3 specifically includes: S31. The Gaussian radial basis function (RBF) is chosen as the basic kernel function, and its expression is: Where, x i x j For each sample point in the power load time series, σ is the kernel width parameter, used to adjust the sensitivity of the Gaussian kernel function to the distance between samples; S32. Based on the time-series dependence of power load, construct a time-series kernel function, specifically including: Centered on a time point, the feature space is divided into m windows T = {W1, W2, ..., W...} m Each window corresponds to a different time neighborhood; Window correction function and modified kernel function The formulas are as follows: Define the timing kernel function as follows: S33. Conduct impact factor analysis and quantification; S3301. Based on the time-series characteristics of the power load, the following key influencing factors are determined: current hourly temperature T. t Average temperature T 3 hours before and after t-3 The average temperature T over the past 24 hours a Monthly highest temperature T max Holiday indicator Hs, linear temperature trend; S3302. Construct a factor function Factor(θ) to quantify the contribution of each factor to the power load. The calculation formula is as follows: Factor(θ)=θ1Trend+θ2(η1T max +η2T t +η3T t-3 +η4T a )+θ3Hs stθ1+θ2+θ3=1, η1+η2+η3+η4=1 Among them, the typical value of the trend coefficient θ1 is 0.75; the typical value of the temperature coefficient θ2 is 0.15; the temperature sub-coefficients η1 = 0.05, η2 = 0.7, η3 = 0.15, η4 = 0.1; and the typical value of the holiday coefficient θ3 is 0.

1.

5. The method according to claim 1, characterized in that: Step 4 specifically includes: S41. Based on the time window weight of the time series kernel function and the feature importance score of the influencing factor, the dataset is divided into information particles with different degrees of mixing. S42. Based on the temporal correlation strength of the information particles, i.e. the sensitivity of the information particles, it is determined by the kernel width parameter σ and the contribution value of the influence factor in the temporal kernel function, and is transferred to different granulation levels in real time. S43. Use fine-grained partitioning for information particles with high time sensitivity and coarse-grained partitioning for information particles with low sensitivity.

6. The method according to claim 1, characterized in that: Step S5 includes: S51. Based on the dynamic multi-level granular dataset obtained in S4, the training set and test set are divided in chronological order, and the data is normalized. S52. Based on the dataset processed in S51, construct a time-sensitive load forecasting model, wherein: the time-series kernel function defined in S3 is used as the model kernel function, and the influence factors quantified in S3 are used as the model input features; the decision function is in the form of: in, For the support vector coefficients, b * For bias terms; S53. Input the test set into the trained model, assign high weights to high-granularity data and low weights to low-granularity data, and use the decision function to predict future power load. The output result is a time-continuous load value sequence.

Images