Short-term power load prediction method

Abstract

The invention relates to the technical field of power load prediction, in particular to a short-term power load prediction method, which comprises the following steps: clustering original power load data by using a Q-RPSO-DBSCAN fusion clustering algorithm; performing secondary modal decomposition on each type of power load data in a clustering result; constructing a power load prediction model based on time sequence component decomposition according to the decomposed data; determining similar day types according to the meteorological data vector of the prediction day and the Euclidean distance of the meteorological data corresponding to the plurality of types of data class centers; inputting the data in the similar day category into a power load prediction model to output a power load initial value; training a full-connection neural network through meteorological data corresponding to the original power load data, and outputting a power load training value of a prediction day; and calculating a power load prediction value according to the power load initial value and the power load training value. According to the invention, the efficiency and accuracy of power load prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load prediction, and particularly relates to a short-term power load prediction method. BACKGROUND

[0002] In the prior art, power load prediction is a core technology in an energy management system, and the goal is to accurately estimate power demand in a future period of time according to historical load data, meteorological information, economic indicators, date types and various other factors.

[0003] Chinese Patent Publication No. CN119377889A discloses a short-term power load prediction method, which includes: obtaining historical load data of a historical period, historical meteorological data and historical time data corresponding to the historical load data, and future meteorological data and future time data of a to-be-predicted period; performing feature analysis according to the historical load data, the historical meteorological data and the historical time data to obtain influence factor features of the historical load data; determining influence factor data from the historical meteorological data, the historical time data, the future meteorological data and the future time data according to the influence factor features; performing modal decomposition on the historical load data to obtain load components of different frequencies; determining a prediction result corresponding to each load component according to a preset time series prediction model, the influence factor data and each load component; and merging and reconstructing each prediction result to obtain power load prediction data of the to-be-predicted period. As can be seen, the short-term power load prediction method has the problem that, due to the sensitivity of the DBSCAN clustering algorithm to the neighborhood radius and the minimum sample number, the load data clustering is inaccurate, and the power load data contains complex high-frequency components such as random fluctuations and low-frequency components such as trends and cycles, which further causes the prediction model to fail to capture key features. SUMMARY

[0004] To this end, the present application provides a short-term power load prediction method to overcome the problem in the prior art that, due to the sensitivity of the DBSCAN clustering algorithm to the neighborhood radius and the minimum sample number, the load data clustering is inaccurate, and the power load data contains complex high-frequency components such as random fluctuations and low-frequency components such as trends and cycles, which further causes the prediction model to fail to capture key features.

[0005] To achieve the above-mentioned purpose, the present application provides a short-term power load prediction method, which includes:

[0006] Step S1, collecting original power load data, and using a Q-RPSO-DBSCAN fusion clustering algorithm to cluster the original power load data to obtain a clustering result of the original power load data and meteorological data corresponding to data class centers of a plurality of classes;

[0007] Step S2, performing secondary modal decomposition on each class of power load data in the clustering result to obtain high-frequency components and low-frequency components of each class of power load data respectively, and outputting the decomposed data according to the number of high-frequency components;

[0008] Step S3, constructing a power load prediction model based on time series component decomposition according to the decomposed data;

[0009] Step S4, determining a similar day category according to a Euclidean distance of meteorological data corresponding to a data class center of a plurality of categories and a meteorological data vector of a prediction day;

[0010] Step S5, inputting data in the similar day category into the power load prediction model to output a power load initial value;

[0011] Step S6, training a full connection neural network through meteorological data corresponding to the original power load data, and inputting meteorological data of a prediction day into the trained full connection neural network to output a power load training value of the prediction day;

[0012] Step S7, calculating a power load prediction value according to the power load initial value and the power load training value.

[0013] Further, the step S1 comprises:

[0014] Step S11, optimizing parameters of a DBSCAN algorithm using a Q-RPSO algorithm to form the Q-RPSO-DBSCAN fusion clustering algorithm;

[0015] Step S12, assuming that the original power load data is divided into a plurality of data classes by the Q-RPSO-DBSCAN fusion clustering algorithm, calculating a within-class error sum of squares of the data classes, and denoted as a first within-class error sum of squares;

[0016] Step S13, defining a data class center of a cth data class in a data class as Load c , and c∈{1,2,3……,C};

[0017] Step S14, calculating a Euclidean distance from each data class center to the rest of the data class centers respectively, sorting all Euclidean distances from small to large, and denoted as a to-be-merged sequence MergeList, and an expression of the to-be-merged sequence MergeList is: ,

[0018] wherein, is a Euclidean distance between each two data class centers;

[0019] Step S15, sequentially traversing each element d in the to-be-merged sequence MergeList, merging the data classes corresponding to the two data class centers corresponding to d to obtain a merged data class;

[0020] Step S16, recalculating the data class centers of the merged data class;

[0021] Step S17, obtaining the within-class error sum of squares of the merged data class, and denoted as a second within-class error sum of squares;

[0022] Step S18, if the second within-class error sum of squares is less than the first within-class error sum of squares, retaining the merged data class, and recalculating the Euclidean distance from each data class center of the merged data class to the data class centers of the remaining data classes to update the to-be-merged sequence MergeList, and repeating steps S15-S18;

[0023] Step S19, if the second within-class error sum of squares is greater than or equal to the first within-class error sum of squares, removing d from the to-be-merged sequence MergeList;

[0024] Step S110, if the to-be-merged sequence MergeList after removal is not an empty set, repeating steps S15-S110;

[0025] Step S111, if the to-be-merged sequence MergeList after removal is an empty set, outputting the clustering result Cls and the data class centers Load c of each class of data;

[0026] Step S112, outputting the clustering result Cls and the data class centers Load c of each class of data corresponding to the original power load data and the meteorological data corresponding to the data class centers of the plurality of classes; c-weather .

[0027] Further, the step S2 comprises:

[0028] Step S21, using the CEEMDAN algorithm to decompose each class of data in the clustering result Cls of the original power load data to obtain a plurality of frequency modal functions, denoted as IMF cf ,

[0029] wherein f∈{1,2,3……,F}, F is the number of modal functions obtained after the CEEMDAN algorithm is used to decompose the cth class of data;

[0030] Step S22, obtaining the modal function as IMFcf The components are divided into high-frequency components and low-frequency components, and the high-frequency components are denoted as... The low-frequency component is denoted as The number of high-frequency components is denoted as nHigh, and the number of low-frequency components is denoted as nLow;

[0031] Step S23, use the VMD algorithm to... The high-frequency components are decomposed into high-frequency decomposed components and low-frequency decomposed components. The high-frequency decomposed components are denoted as... The number of high-frequency decomposition components is denoted as The low-frequency decomposition component is denoted as The number of low-frequency decomposition components is denoted as ;

[0032] Step S24, if If the result is less than nHigh, then the decomposition result of the VMD algorithm is retained;

[0033] Step S25, if If the result is greater than or equal to nHigh, then retain the decomposition result of the CEEMDAN algorithm;

[0034] Step S26, the decomposed data output from the retained decomposition results is denoted as... .

[0035] Further, step S3 includes:

[0036] Step S31, use the formula for calculating the smoothing filter value to... Each of them The data is smoothed and filtered to obtain the filtered data.

[0037] Step S32: Obtain the trend prediction component trend(t) of the time series sequence at time t in the filtered data;

[0038] Step S33, for Each of them Data Computing The periodic component Period(t) of the time series is predicted;

[0039] Step S34, for Each of them Data Computing The seasonal component Seasonal(t) is predicted from the time series sequence;

[0040] Step S35, for Each of them Data Computing Predicting the nonstationary component (NonStationary(t)) of a time series;

[0041] Step S36, for Each of them Data Computing The random perturbation noise (Noise(t)) for time series prediction is calculated using the following formula:

[0042] ,

[0043] Where rand is a random number on [0, 1] that follows a uniform distribution, A1 is the average value of the power load data sequence, and A2 is the minimum value of the power load data sequence;

[0044] Step S37, based on trend(t), period(t), nonstationary(t), and noise(t), establish the power load prediction model at time t as follows:

[0045] .

[0046] Further, step S31 includes:

[0047] Step S311, for The smoothed filter value of the data at time t is expressed as: , where ns < t;

[0048] Step S312: If there are fewer than ns time series data before time t, then no smoothing filtering is performed.

[0049] Step S313: Calculate the weather data at time t. t The vector and the corresponding meteorological data vectors at times t-1, t-2, ..., t-ns are respectively weather t-1 weather t-2 ...weather t-ns The Euclidean distances are denoted as dl1, dl2, ..., dl ns And q1 to qns are from Sure;

[0050] Step S314: Perform polynomial fitting on the filtered data to fit it as follows: ,

[0051] Where t represents the time of day when the electricity load is highest, and t is an integer. These are the parameters for the polynomial fitting.

[0052] Step S315: Calculate the trend prediction component of the time series at time t, trend(t). The formula for calculating trend(t) is as follows: , where α is a constant.

[0053] Further, step S33 includes:

[0054] Step S331: Calculate the slope of each pair of adjacent data points for the first ns time steps before time t, and denote the slope of data point x. t-1 x t-2 x t-3 ... x t-ns+1 The corresponding slopes are k t-1 k t-2 k t-3 ... k t-ns+1 ;

[0055] Step S332: Obtain the slope sequence k using a single exponential smoothing method. t-ns+1 ... k t-3 k t-2 k t-1 The slope value k at the next moment t The formula for calculating the first exponential smoothing method is:

[0056] ,

[0057] Among them, ks t-1 This is the predicted slope value from the previous moment. It is a constant;

[0058] Step S333, for the slope sequence k t-ns+1 ... k t-3 k t-2 k t-1 k t Normalization is performed to obtain normalized slope sequence data. The normalization calculation formula is:

[0059] ,

[0060] Where, k i Let k be the i-th value in the slope sequence, min(k) be the minimum value in the slope sequence, and max(k) be the maximum value in the slope sequence.

[0061] Step S334, based on the normalized slope sequence data The expression for determining the predicted periodic component Period(t) of the time series is as follows:

[0062] ,

[0063] Where M is a constant, T is the period, and the period is 365.

[0064] Further, step S34 includes:

[0065] Step S341: For time t, denote the power load data sequence of the previous ns time steps as Load. cur Let Load be the sequence of power load data for the first ns moments of the same period last year as time t. pre ;

[0066] Step S342: If data from the previous year at the same time as time t exists, then denote the meteorological data vector at time t as Load. cur_weather The meteorological data vector for the same period of the previous year at time t is Load. pre_weather ;

[0067] Step S343, calculate Load cur_weather and Load pre_weather The cosine similarity value Sim1 is used to calculate Load. cur and Load pre The cosine similarity Sim2 is calculated using the following formula:

[0068] ,

[0069] Step S344: When Sim1 is less than the preset cosine similarity value and Sim2 is less than the preset cosine similarity value, Seasonal(t) is 0.

[0070] Step S345: Under the condition that Sim1 is less than a preset cosine similarity value and Sim2 is less than a preset cosine similarity value, ,in, This represents the seasonality impact coefficient. This refers to the electricity load data for the same period at time t in the previous year;

[0071] Step S346: If there is no data from the same period of the previous year as time t, then Seasonal(t) is 0.

[0072] Further, step S35 includes:

[0073] Step S351: Concatenate the meteorological data vector with the power load data from the ns time points prior to time t. ;

[0074] Step S352, will The input is fed into a single-layer BP neural network and the output is Non-Stationary(t) at time t.

[0075] Further, step S4 includes:

[0076] Step S41, denote the meteorological data vector for the predicted day as Load. forcast_weather Calculate Load forcast_weather Euclidean distance between the meteorological data vectors of the data center of each data class in Cls and the data center of each data class in Cls;

[0077] Step S42: The data category corresponding to the data class center with the smallest Euclidean distance is taken as the similarity category for the prediction date, and the similarity category is denoted as s, and the data class center of the similarity category is Load. s , The component data of the mode function corresponding to similar categories are denoted as .

[0078] Furthermore, the formula for calculating the predicted power load is as follows:

[0079] ,

[0080] Among them, ForcastLoad final The ForcastLoad is the predicted power load value. seq ForcastLoad is the initial value of the power load. FCNN This refers to the training value for the power load.

[0081] Compared with existing technologies, the beneficial effects of this invention are as follows: The method of this invention uses the Q-RPSO-DBSCAN fusion clustering algorithm to cluster the original power load data, overcoming the problem that the traditional DBSCAN clustering algorithm is highly sensitive to the neighborhood radius and minimum sample number parameters, resulting in unstable clustering results and inaccurate category classification; by introducing a quantum behavior-improved recursive particle swarm optimization algorithm to automatically optimize the key parameters of DBSCAN, the adaptability of the clustering process is improved, ensuring that load data under different feature patterns can be reasonably classified, thereby improving the data representativeness and classification accuracy of subsequent modeling; by performing secondary mode decomposition, efficient separation and reconstruction of high-frequency random fluctuation components and low-frequency trend periodic components in complex load signals are achieved; by judging whether the number of high-frequency components after VMD decomposition exceeds a preset threshold to select the optimal decomposition path, information distortion caused by over-decomposition or under-decomposition is avoided, and the stability of mode decomposition is improved.

[0082] Furthermore, the method described in this invention constructs a power load forecasting model based on time series component decomposition, decomposing the filtered load data into a trend term (t), a period term (t), a seasonal term (t), a nonstationary term (t), and a noise term (t). By introducing a meteorological data-driven trend smoothing mechanism and an exponential smoothing-based slope extrapolation method, the model's response to sudden load changes is enhanced, and the accuracy of capturing short-term load change trends is improved. The CEEMDAN algorithm decomposes each type of data into multiple mode functions, distinguishing between high-frequency and low-frequency components. Subsequently, the VMD algorithm performs a secondary decomposition on the high-frequency components, further separating the high-frequency and low-frequency decomposed components. The secondary mode decomposition decomposes the original data into multiple relatively stationary sub-components, reducing the interference of nonstationarity on the forecast. VMD further decomposes high-frequency components to extract more refined stationary components, thereby effectively improving data quality and the predictability of model input. By comparing the number of high-frequency components, it decides whether to retain CEEMDAN or VMD results, avoiding over-decomposition or information loss, making the decomposed data more suitable for model building. Especially in high-fluctuation load scenarios, the addition of stationary components improves prediction accuracy.

[0083] Furthermore, the method of the present invention, when determining the similar day category to which the prediction day belongs, achieves effective matching of meteorological load relationships based on the Euclidean distance between the meteorological conditions of the prediction day and the meteorological data corresponding to the center of each data category. This improves the correlation between the input sample and historical typical scenarios, enabling the prediction model to more accurately call upon the historical load evolution patterns closest to the current environment, thereby enhancing the model's scenario adaptability. The selection of similar days, through meteorological data vector matching, aligns the prediction day with historical data categories, ensuring the correlation of the model input, reducing prediction bias caused by meteorological factors such as temperature and humidity, and improving the reliability of short-term predictions.

[0084] Furthermore, the method described in this invention independently models the relationship between meteorological factors and load response using a fully connected neural network to obtain power load training values ​​that depend solely on meteorological variables. These training values ​​are then weighted and fused with initial power load values ​​based on similar day categories to form the final prediction output. This dual-channel prediction architecture retains the trend guidance based on historical load patterns while introducing the correction capability of meteorological sensitivity modeling, thereby improving the accuracy of the model under extreme weather or load abrupt changes. Attached Figure Description

[0085] Figure 1 This is an overall flowchart of the short-term power load forecasting method according to an embodiment of the present invention;

[0086] Figure 2This is a flowchart of step S1 of the short-term power load forecasting method according to an embodiment of the present invention;

[0087] Figure 3 This is an overall logic block diagram of the short-term power load forecasting method according to an embodiment of the present invention. Detailed Implementation

[0088] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0089] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0090] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0091] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0092] Please see Figure 1 , Figure 2 , Figure 3 The figures shown are the overall flowchart, the flowchart of step S1, and the overall logic block diagram of the short-term power load forecasting method according to an embodiment of the present invention. The present invention provides a short-term power load forecasting method, comprising:

[0093] Step S1: Collect raw power load data and use the Q-RPSO-DBSCAN fusion clustering algorithm to cluster the raw power load data to obtain the clustering results of the raw power load data and the meteorological data corresponding to the data center of several classes;

[0094] Step S2: Perform secondary mode decomposition on each type of power load data in the clustering results to obtain the high-frequency component and low-frequency component of each type of power load data respectively, and output the decomposed data according to the number of the high-frequency component.

[0095] Step S3: Construct an electricity load prediction model based on time series component decomposition based on the decomposed data;

[0096] Step S4: Determine the similar day category based on the meteorological data vector of the predicted day and the Euclidean distance of the meteorological data corresponding to the data class centers of several classes;

[0097] Step S5: Input the data from the similar day categories into the power load prediction model to output the initial power load value;

[0098] Step S6: Train the fully connected neural network using the meteorological data corresponding to the original power load data, and input the meteorological data of the prediction day into the trained fully connected neural network to output the power load training value of the prediction day.

[0099] Step S7: Calculate the predicted power load value based on the initial power load value and the training power load value.

[0100] The power load data used in this embodiment is the current power load value recorded every 15 minutes, so there are a total of 96 data points for the power load curve in a day; the meteorological data vector corresponding to the recorded power load is the four data points of temperature (Temp), humidity (Hum), rainfall (Rain), and wind (Wind) at the time the load value is recorded.

[0101] Specifically, Q-RPSO is a quadratic reconstruction particle swarm optimization algorithm, and the position update formula of the Q-RPSO algorithm is:

[0102] ,

[0103] Where, x t+1 The particle position at the next moment; c is one-fifth to one-third of the maximum value of the particle position; r is a random number in the interval [-1, 1]; t is the current iteration number of the Q-RPSO algorithm; T is the maximum iteration number of the Q-RPSO algorithm; a is a random number in the interval [0, 1]; p i,j (t) represents the historical best position of the i-th particle; p g,j (t) represents the position of the globally optimal particle.

[0104] Specifically,

[0105] In implementation, the density-based spatial clustering algorithm DBSCAN (Density-Based Spatial Clustering of Applications with Noise) defines each data class as the largest set of density-connected points. It determines the neighborhood of a data point using a parameter scan radius (Eps) and checks if the number of data points within that neighborhood is greater than or equal to the minimum number of contained points (MinPts). If so, the data within that neighborhood is considered a single data class, called a cluster. The DBSCAN algorithm forms clusters by recursively modifying the neighborhood of core objects. It can discover clusters of arbitrary shapes in noisy data and has the ability to remove noise. However, the clustering performance of the DBSCAN algorithm is significantly affected by the parameters Eps and MinPts, and currently, the scan radius and minimum number of contained points are mostly determined experimentally.

[0106] Specifically, the encoding of the Eps and minPts parameters is optimized using the Q-RPSO algorithm as follows: .

[0107] Specifically, the fitness function of the Q-RPSO-DBSCAN algorithm is the sum of squared errors within the class (SSE), and the optimization objective is to minimize the SSE. The formula for calculating the SSE is:

[0108]

[0109] Where M is the number of data items contained in a certain type of data, Load n For the nth power load data within this data class, Load c This serves as the class center for this type of power load data.

[0110] In implementation, the method of this invention uses the Q-RPSO-DBSCAN fusion clustering algorithm to cluster the original power load data, overcoming the problem that the traditional DBSCAN clustering algorithm is highly sensitive to the neighborhood radius and minimum sample number parameters, resulting in unstable clustering results and inaccurate category classification. By introducing a quantum behavior-improved recursive particle swarm optimization algorithm to automatically optimize the key parameters of DBSCAN, the adaptability of the clustering process is improved, ensuring that load data under different feature patterns can be reasonably classified, thereby improving the data representativeness and classification accuracy of subsequent modeling. By performing secondary mode decomposition, efficient separation and reconstruction of high-frequency random fluctuation components and low-frequency trend periodic components in complex load signals are achieved. By judging whether the number of high-frequency components after VMD decomposition exceeds a preset threshold to select the optimal decomposition path, information distortion caused by over-decomposition or under-decomposition is avoided, and the stability of mode decomposition is improved.

[0111] Specifically, step S1 includes:

[0112] Step S11: Optimize the parameters of the DBSCAN algorithm using the Q-RPSO algorithm to form the Q-RPSO-DBSCAN fusion clustering algorithm;

[0113] Step S12: Suppose that the original power load data is divided into several data classes by the Q-RPSO-DBSCAN fusion clustering algorithm, calculate the sum of squared errors within each data class, and denot it as the first sum of squared errors within each data class;

[0114] Step S13: Define the data class center of the c-th data in the data class as Load. c And c∈{1,2,3……,C};

[0115] Step S14: Calculate the Euclidean distance from each data class center to the centers of the other data classes, sort all Euclidean distances from smallest to largest, and denote them as the merged sequence MergeList. The expression for the merged sequence MergeList is: ,

[0116] in, The Euclidean distance between the centers of every two data classes;

[0117] Step S15: Iterate through each element d in the MergeList sequence to be merged, and merge the data classes corresponding to the two data class centers corresponding to d to obtain the merged data class;

[0118] Step S16: Recalculate the data class center of the merged data class;

[0119] Step S17: Obtain the sum of squared errors within the merged data class and record it as the second sum of squared errors within the class;

[0120] Step S18: If the sum of squared errors within the second class is less than the sum of squared errors within the first class, then retain the merged data class and recalculate the Euclidean distance from the center of each data class to the centers of the other data classes to update the merged sequence MergeList. Repeat steps S15-S18.

[0121] Step S19: If the sum of squared errors within the second class is greater than or equal to the sum of squared errors within the first class, then remove d from the MergeList of sequences to be merged.

[0122] Step S110: If the MergeList sequence to be merged after removal is not an empty set, then repeat steps S15-S110.

[0123] Step S111: If the MergeList sequence to be merged after removal is an empty set, then output the clustering result Cls and the data cluster center Load for each data class. c ;

[0124] Step S111: Output the clustering results Cls and the cluster centers Load for each cluster. c The corresponding clustering results Cls of the original power load data and the meteorological data Load corresponding to the data center of several classes. c-weather .

[0125] Specifically, step S2 includes:

[0126] Step S21: For each cluster of data in the original power load data Cls, the CEEMDAN algorithm is used to decompose the data to obtain mode functions of several frequencies. Let the mode function of the c-th data in Cls after CEEMDAN decomposition be IMF. cf ,

[0127] Where f∈{1,2,3……,F}, and F is the number of modal functions obtained after the c-th class of data is decomposed by the CEEMDAN algorithm;

[0128] Step S22, set the mode function to IMF cf The components are divided into high-frequency components and low-frequency components, and the high-frequency components are denoted as... The low-frequency component is denoted as The number of high-frequency components is denoted as nHigh, and the number of low-frequency components is denoted as nLow;

[0129] Step S23, use the VMD algorithm to... The high-frequency components are decomposed into high-frequency decomposed components and low-frequency decomposed components. The high-frequency decomposed components are denoted as... The number of high-frequency decomposition components is denoted as The low-frequency decomposition component is denoted as The number of low-frequency decomposition components is denoted as ;

[0130] Step S24, if If the result is less than nHigh, then the decomposition result of the VMD algorithm is retained;

[0131] Step S25, if If the result is greater than or equal to nHigh, then retain the decomposition result of the CEEMDAN algorithm;

[0132] Step S26, the decomposed data output from the retained decomposition results is denoted as... .

[0133] Specifically, step S3 includes:

[0134] Step S31, use the formula for calculating the smoothing filter value to... Each of them The data is smoothed and filtered to obtain the filtered data.

[0135] Step S32: Obtain the trend prediction component trend(t) of the time series sequence at time t in the filtered data;

[0136] Step S33, for Each of them Data Computing The periodic component Period(t) of the time series is predicted;

[0137] Step S34, for Each of them Data Computing The seasonal component Seasonal(t) is predicted from the time series sequence;

[0138] Step S35, for Each of them Data Computing Predicting the nonstationary component (NonStationary(t)) of a time series;

[0139] Step S36, for Each of them Data Computing The random perturbation noise (Noise(t)) for time series prediction is calculated using the following formula:

[0140] ,

[0141] Where rand is a random number on [0, 1] that follows a uniform distribution, A1 is the average value of the power load data sequence, and A2 is the minimum value of the power load data sequence;

[0142] Step S37, based on trend(t), period(t), nonstationary(t), and noise(t), establish the power load prediction model at time t as follows:

[0143] .

[0144] Specifically, high-frequency and low-frequency components are distinguished based on signal frequency. Signals with a frequency greater than or equal to 300Hz are classified as high-frequency signals, while those with a frequency less than 300Hz are classified as low-frequency signals.

[0145] In implementation, the method of this invention constructs a power load forecasting model based on time series component decomposition, decomposing the filtered load data into a trend term (t), a period term (t), a seasonal term (t), a nonstationary term (t), and a noise term (t). By introducing a meteorological data-driven trend smoothing mechanism and an exponential smoothing-based slope extrapolation method, the model's response to sudden load changes is enhanced, and the accuracy of capturing short-term load change trends is improved. The CEEMDAN algorithm decomposes each type of data into multiple mode functions, distinguishing between high-frequency and low-frequency components. Subsequently, the VMD algorithm performs a secondary decomposition on the high-frequency components, further separating the high-frequency and low-frequency decomposed components. The secondary mode decomposition decomposes the original data into multiple relatively stationary sub-components, reducing the interference of nonstationarity on the forecast. VMD further decomposes high-frequency components to extract more refined stationary components, thereby effectively improving data quality and the predictability of model input. By comparing the number of high-frequency components, it decides whether to retain CEEMDAN or VMD results, avoiding over-decomposition or information loss, making the decomposed data more suitable for model building. Especially in high-fluctuation load scenarios, the addition of stationary components improves prediction accuracy.

[0146] Specifically, step S31 includes:

[0147] Step S311, for The smoothed filter value of the data at time t is expressed as: , where ns < t;

[0148] Step S312: If there are fewer than ns time series data before time t, then no smoothing filtering is performed.

[0149] Step S313: Calculate the weather data at time t. t The vector and the corresponding meteorological data vectors at times t-1, t-2, ..., t-ns are respectively weather t-1 weather t-2 ...weather t-ns The Euclidean distances are denoted as dl1, dl2, ..., dl ns And q1 to qns are from Sure;

[0150] Step S314: Perform polynomial fitting on the filtered data to fit it as follows: ,

[0151] Where t represents the time of day when the electricity load is highest, and t is an integer. These are the parameters for the polynomial fitting.

[0152] Step S315: Calculate the trend prediction component of the time series at time t, trend(t). The formula for calculating trend(t) is as follows: , where α is a constant.

[0153] Specifically, the value of time t ranges from [1, 96].

[0154] Specifically, Solve using the least squares method.

[0155] Specifically, the value of α is 0.8.

[0156] Specifically, step S33 includes:

[0157] Step S331: Calculate the slope of each pair of adjacent data points for the first ns time steps before time t, and denote the slope of data point x. t-1 x t-2 x t-3 ... x t-ns+1 The corresponding slopes are k t-1 k t-2 k t-3 ... k t-ns+1 ;

[0158] Step S332: Obtain the slope sequence k using a single exponential smoothing method. t-ns+1 ... k t-3 k t-2 k t-1 The slope value k at the next moment t The formula for calculating the first exponential smoothing method is:

[0159] ,

[0160] Among them, ks t-1 This is the predicted slope value from the previous moment. It is a constant;

[0161] Step S333, for the slope sequence k t-ns+1 ... k t-3 k t-2 k t-1 k t Normalization is performed to obtain normalized slope sequence data. The normalization calculation formula is:

[0162] ,

[0163] Where, k iLet k be the i-th value in the slope sequence, min(k) be the minimum value in the slope sequence, and max(k) be the maximum value in the slope sequence.

[0164] Step S334, based on the normalized slope sequence data The expression for determining the predicted periodic component Period(t) of the time series is as follows:

[0165] ,

[0166] Where M is a constant, T is the period, and the period is 365.

[0167] Specifically, The value is 0.6.

[0168] Specifically, the value of M is 3.

[0169] Specifically, step S34 includes:

[0170] Step S341: For time t, denote the power load data sequence of the previous ns time steps as Load. cur Let Load be the sequence of power load data for the first ns moments of the same period last year as time t. pre ;

[0171] Step S342: If data from the previous year at the same time as time t exists, then denote the meteorological data vector at time t as Load. cur_weather The meteorological data vector for the same period of the previous year at time t is Load. pre_weather ;

[0172] Step S343, calculate Load cur_weather and Load pre_weather The cosine similarity value Sim1 is used to calculate Load. cur and Load pre The cosine similarity Sim2 is calculated using the following formula:

[0173] ,

[0174] Step S344: When Sim1 is less than the preset cosine similarity value and Sim2 is less than the preset cosine similarity value, Seasonal(t) is 0.

[0175] Step S345: Under the condition that Sim1 is less than a preset cosine similarity value and Sim2 is less than a preset cosine similarity value, ,in, This represents the seasonality impact coefficient. This refers to the electricity load data for the same period at time t in the previous year;

[0176] Step S346: If there is no data from the same period of the previous year as time t, then Seasonal(t) is 0.

[0177] Specifically, the first ns time steps refer to the first 96 time steps.

[0178] Specifically, The value is 0.35.

[0179] Specifically, if both Sim1 and Sim2 are less than the preset cosine similarity value, it indicates that there is no seasonal correlation between the two times; if both Sim1 and Sim2 are greater than the preset cosine similarity value, it indicates that there is a seasonal correlation between the two times.

[0180] Specifically, the preset cosine similarity value is 0.6.

[0181] Specifically, step S35 includes:

[0182] Step S351: Concatenate the meteorological data vector with the power load data from the ns time points prior to time t. ;

[0183] Step S352, will The input is fed into a single-layer BP neural network and the output is Non-Stationary(t) at time t.

[0184] Specifically, the number of hidden layer neurons in a single-layer BP neural network is set to 10, and the number of output neurons is 1.

[0185] Specifically, step S4 includes:

[0186] Step S41, denote the meteorological data vector for the predicted day as Load. forcast_weather Calculate Load forcast_weather Euclidean distance between the meteorological data vectors of the data center of each data class in Cls and the data center of each data class in Cls;

[0187] Step S42: The data category corresponding to the data class center with the smallest Euclidean distance is taken as the similarity category for the prediction date, and the similarity category is denoted as s, and the data class center of the similarity category is Load. s , The component data of the mode function corresponding to similar categories are denoted as .

[0188] In practice, the method of this invention, when determining the similar day category to which the prediction day belongs, achieves effective matching of meteorological load relationships based on the Euclidean distance between the meteorological conditions of the prediction day and the meteorological data corresponding to the center of each data category. This improves the correlation between the input sample and historical typical scenarios, enabling the prediction model to more accurately call upon the historical load evolution patterns closest to the current environment, thus enhancing the model's scenario adaptability. The selection of similar days, through meteorological data vector matching, aligns the prediction day with historical data categories, ensuring the correlation of the model input, reducing prediction bias caused by meteorological factors such as temperature and humidity, and improving the reliability of short-term predictions.

[0189] Specifically, step S5 includes:

[0190] Will The initial value of the power load is input into the power load prediction model to calculate the initial value of the power load, and the initial value of the power load is denoted as ForcastLoad. seq .

[0191] Specifically, step S6 involves inputting the historical gas phase data into the power load prediction model to obtain the power load reference value at the corresponding time; including:

[0192] Step S61: Use the meteorological data corresponding to the original power load data as input and the power load value as output to train the fully connected neural network;

[0193] Step S62: Input the meteorological data for the predicted day into the trained fully connected neural network to output the power load training value, and record the power load training value as ForcastLoad. FCNN .

[0194] Specifically, the number of hidden layers in the fully connected neural network is set to 2, the number of neurons in the first and second hidden layers is set to 32, and the learning rate is 1.0.

[0195] In practice, the method described in this invention independently models the relationship between meteorological factors and load response through a fully connected neural network, obtains power load training values ​​that depend only on meteorological variables, and then weights and fuses them with initial power load values ​​based on similar day categories to form the final prediction output. This dual-channel prediction architecture retains the trend guidance based on historical load patterns while introducing the correction capability of meteorological sensitivity modeling, thereby improving the accuracy of the model under extreme weather or load change conditions.

[0196] Specifically, the formula for calculating the predicted power load is as follows:

[0197] ,

[0198] Among them, ForcastLoad final The ForcastLoad is the predicted power load value. seq ForcastLoad is the initial value of the power load. FCNN This refers to the training value for the power load.

[0199] In this embodiment, constant α, constant β, constant M, and seasonality influence coefficient are used. This is the result of the experiment in this embodiment.

[0200] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A short-term electric power load forecasting method characterized by, The method comprises the following steps: Step S1, collecting original power load data and clustering the original power load data to obtain clustering results of the original power load data and meteorological data corresponding to data class centers of several classes by using a Q-RPSO-DBSCAN fusion clustering algorithm; Step S2, performing secondary modal decomposition on each class of power load data in the clustering results to obtain high-frequency components and low-frequency components of each class of power load data, respectively, and outputting the decomposed data according to the number of high-frequency components; Step S3, constructing a power load prediction model based on time sequence component decomposition according to the decomposed data; Step S4, determining a similar day category according to a meteorological data vector of a prediction day and the Euclidean distance of the meteorological data corresponding to the data class centers of the several classes; Step S5, inputting data in the similar day category into the power load prediction model to output an initial value of power load; Step S6, training a full connection neural network through the meteorological data corresponding to the original power load data, and inputting the meteorological data of the prediction day into the trained full connection neural network to output a training value of power load of the prediction day; Step S7, calculating a power load prediction value according to the initial value of power load and the training value of power load.

2. The short-term power load forecasting method according to claim 1, characterized in that, The step S1 comprises: Step S11, optimizing parameters of a DBSCAN algorithm by using a Q-RPSO algorithm to form the Q-RPSO-DBSCAN fusion clustering algorithm; Step S12, calculating the within-class sum of squares of errors of the data classes, and denoted as a first within-class sum of squares of errors, by dividing the original power load data into several data classes by using the Q-RPSO-DBSCAN fusion clustering algorithm; Step S13, defining the data class center of the cth data in the data class as Load c , and c e {1, 2, 3, …, C}; Step S14, respectively calculate the Euclidean distance from each of the data class center to the rest of the data class center, sort all the Euclidean distance from small to large, and record as the to-be-merged sequence MergeList, the expression of the to-be-merged sequence MergeList is: , wherein, is the Euclidean distance between each two of the data clusters. Step S15, sequentially traversing each element d in the to-be-merged sequence MergeList, merging the data classes corresponding to the two data class centers corresponding to d to obtain a merged data class; Step S16, recalculating the data class centers of the merged data class; Step S17, calculating the within-class sum of squares of errors of the merged data class, and denoted as a second within-class sum of squares of errors; Step S18, if the second within-class sum of squares of errors is less than the first within-class sum of squares of errors, retaining the merged data class, and recalculating the Euclidean distances from each data class center of the merged data class to the data class centers of the remaining data classes to update the to-be-merged sequence MergeList, and repeating steps S15-S18; Step S19, if the second within-class sum of squares of errors is greater than or equal to the first within-class sum of squares of errors, removing d from the to-be-merged sequence MergeList; Step S110, if the to-be-merged sequence MergeList after removal is not an empty set, repeating steps S15-S110; Step S111, if the removed MergeList is empty set, output the clustering result Cls and the data class center Load of each class data c ; Step S112, outputting the clustering result Cls and the data class center Load of each class data c The clustering result Cls and the data class center of the plurality of classes corresponding to the original power load data c-weather .

3. The short-term electric power load forecasting method according to claim 2, characterized by, The step S2 comprises: Step S21, using the CEEMDAN algorithm to decompose each class data of the clustering result Cls of the original power load data to obtain a plurality of frequency mode functions, and recording the mode function of the cth class data in Cls after CEEMDAN decomposition as IMF cf , Wherein, f∈{1,2,3……,F}, F is the number of modal functions obtained after the CEEMDAN algorithm is performed on the cth data; Step S22, the modal function is IMF cf The high frequency component is denoted as The low frequency component is denoted as The number of high frequency components is denoted as nHigh, and the number of low frequency components is denoted as nLow. Step S23, use the VMD algorithm to... The high-frequency components are decomposed into high-frequency decomposed components and low-frequency decomposed components. The high-frequency decomposed components are denoted as... The number of high-frequency decomposition components is denoted as The low-frequency decomposition component is denoted as The number of low-frequency decomposition components is denoted as ; Step S24, if less than nHigh, the decomposition result of the VMD algorithm is reserved; Step S25, if greater than or equal to nHigh, the decomposition result of the CEEMDAN algorithm is reserved; Step S26, the remaining decomposition result output of the post-decomposition data is denoted as .

4. The short-term electric power load forecasting method according to claim 3, characterized by, The step S3 comprises: Step S31, using the calculation formula of the smoothing filter value to perform smoothing filter processing on each of the data in the set of data to obtain filtered data. data to obtain filtered data;​ Step S32, obtaining a trend prediction component trend(t) of the time sequence at time t in the filtered data; Step S33, for Each of them Data Computing Predict the periodic component Period(t) of the time series sequence; Step S34, for each of the time series data, a seasonal component Seasonal(t) is computed ​​ Step S35, for each of the data data computes a time series prediction of the non-stationary component NonStationary(t);​ Step S36, for Each of them Data Computing The random perturbation noise (Noise(t)) for time series prediction is calculated using the following formula: ， Wherein, rand is a random number obeying uniform distribution on [0, 1], A1 is the average value of the power load data sequence, and A2 is the minimum value of the power load data sequence; Step S37, establishing the power load prediction model at time t according to trend(t), Period(t), NonStationary(t) and Noise(t) as follows: 。 5. The short-term electric power load forecasting method according to claim 4, characterized by, The step S31 comprises: Step S311, for The smoothed filtered value of the data at time t is expressed as where ns < t. Step S312, if the time sequence data before time t is less than ns, no smoothing filtering processing is taken; Step S313, respectively calculate the weather data weather t Vector and t-1, t-2,..., t-ns time corresponding to the weather data vector respectively weather t-1 , weather t-2 ,..., weather t-ns Euclidean distance, and recorded as dl1, dl2,...., dl ns , and q1 to qns are determined by ; Step S314, polynomial fitting is performed on the filtered data, fitting is , where t is the power load time of the day, and t is an integer, is a polynomial fit parameter; Step S315, the trend prediction component trend(t) of the time sequence sequence at time t is calculated, and the calculation formula of trend(t) is wherein, a is a constant.

6. The short-term electric power load forecasting method according to claim 5, characterized by, The step S33 comprises: Step S331, calculate the slope of each two adjacent data points of the previous ns time points at time t, and record the data points x t-1 , x t-2 , x t-3 , …, x t-ns+1 , corresponding slopes are k t-1 , k t-2 , k t-3 , …, k t-ns+1 ; Step S332, get the slope sequence k by one exponential smoothing method t-ns+1 , …, k t-3 , k t-2 , k t-1 The slope value k of the next moment t The calculation formula of one exponential smoothing method is: ， where ks t-1 is the slope prediction value of the previous time instant, is a constant; Step S333, normalizing the slope sequence k t-ns+1 , …, k t-3 , k t-2 , k t-1 , k t to obtain normalized slope sequence data The normalization calculation formula is: ， where k i is the i-th value in the slope sequence, min(k) is the minimum value in the slope sequence; max(k) is the maximum value in the slope sequence; Step S334, according to the normalization slope sequence data The expression of the timing sequence prediction period component Period(t) is determined as follows: ， Wherein, M is a constant, and T is a period, and the period takes 365.

7. The short-term electric power load forecasting method according to claim 6, characterized by, The step S34 comprises: Step S341, for t time, record its previous ns time power load data sequence as Load cur , record the previous ns time power load data sequence of the same period as t time last year as Load pre ; Step S342, if there is data of the same period of the previous year as the time t, record the weather data vector of the time t as Load cur_weather , the weather data vector of the same period of the previous year as the time t is Load pre_weather ; Step S343, calculate the cosine similarity value Sim1 of Load cur_weather and Load pre_weather Step S344, calculate the cosine similarity Sim2 of Load cur and Load pre The formula for calculating the cosine similarity value is: ， Step S344, Seasonal(t) is 0 under the condition that Sim1 is less than the preset cosine similarity value and Sim2 is less than the preset cosine similarity value; Step S345, under the condition that Sim1 is not less than the preset cosine similarity value and Sim2 is not less than the preset cosine similarity value, the process proceeds to step S346. wherein, is a seasonal influence coefficient; is the power load data of the same period of the previous year and the t time. Step S346, if there is no data of the same period of the previous year and time t, Seasonal(t) is 0.

8. The short-term electric power load forecasting method according to claim 7, characterized by, The step S35 comprises: Step S351, splicing the meteorological data vector with the power load data of ns time points before the t time point as ; Step S352, the Non-Stationary(t) is input to the single-layer BP neural network and output at time t. input to the single-layer BP neural network and output at time t.

9. The short-term electric power load forecasting method according to claim 8, characterized by, The step S4 comprises: Step S41, let the weather data vector of the prediction day be Load forcast_weather , calculate the Euclidean distance between Load forcast_weather and the weather data vector of each class center of Cls; Step S42, taking the data category corresponding to the data class center with the minimum Euclidean distance as the similar category of the predicted day, recording the similar category as s, and the data class center of the similar category as Load s , , recording the component data of the modal function corresponding to the similar category as .

10. The short-term power load forecasting method according to claim 9, characterized in that, The calculation formula of the power load prediction value is as follows: ， wherein, ForcastLoad final is the power load forecast value, ForcastLoad seq is the power load initial value, ForcastLoad FCNN is the power load training value.

Images