Short-term power load prediction method based on daily load characteristic clustering

Abstract

The invention discloses a daily load characteristic clustering-based short-term power load prediction method. The method comprises the following steps of 1, acquiring power load original data; 2, carrying out standardization processing on the original data; step 3, carrying out daily load characteristic classification processing; step 4, performing causal feature discovery based on a PC (Personal Computer) algorithm; step 5, inputting a PC-TCN prediction model based on daily load characteristic classification; and step 6, measuring and outputting a prediction result. Carrying out data dimension reduction by measuring and calculating daily load characteristics, and replacing original date characteristics with obtained clustering results; meanwhile, based on a causal discovery algorithm, the influence index set is screened, and extraction of causal features is enhanced, so that the operation efficiency and precision are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to a short-term power load prediction method based on daily load characteristic clustering. BACKGROUND

[0002] In the field of power load prediction technology, the prediction method based on neural network model develops rapidly, and various prediction models for optimizing model accuracy emerge. However, there are still related problems in short-term power load prediction. In the time dimension, the load curve of different days is affected by the specific conditions of the day, showing different load characteristics. This inter-day difference reduces the effectiveness of the date index as a classification feature. To address this issue, clustering methods are often used to identify and classify historical daily load curves. However, when faced with high-dimensional daily load curve data, traditional clustering methods are prone to dimension disaster, resulting in unstable clustering results, increased computational complexity, and other problems. At the same time, short-term power load prediction also faces the problem of time-varying indicators. The prediction involves various complex factors such as economic conditions, weather parameters, and holiday types, which are quantified into numerous impact indicators. If the time prior influence is further considered, the dimension of the impact factors will double. At different prediction points, the influence of each factor on the load value varies significantly. This dynamic change means that not all collected impact factors have strong relevance to all time prediction targets, resulting in redundant or weakly correlated indicators in the feature set, which not only increases the complexity of the model but also interferes with the extraction of key features, ultimately affecting the accuracy and computational efficiency of the prediction. SUMMARY

[0003] In view of the above situation, in order to solve the above technical problems such as unstable clustering results, time-varying and redundant indicators in the model in short-term load prediction, the present application proposes a prediction method based on daily load characteristic clustering. The daily load characteristics are calculated to reduce the dimension of the data, and the obtained clustering results are used to replace the original date features. At the same time, based on the causal discovery algorithm, the impact indicator set is filtered to enhance the extraction of causal features, thereby improving the computational efficiency and accuracy.

[0004] A short-term power load prediction method based on daily load characteristic clustering, comprising the following steps:

[0005] Step 1: Obtain the original power load data;

[0006] Step 2: Standardize the original data;

[0007] Step 3: Perform daily load characteristic classification processing;

[0008] Step 4: Causal feature discovery based on PC algorithm;

[0009] Step 5: Input the PC-TCN prediction model based on daily load characteristics classification;

[0010] Step 6: Measure and output the prediction results.

[0011] Step 1: Obtain raw power load data, including:

[0012] The daily 24-hour electricity load dataset for a certain time period within a region is denoted as L=[L1,L2,...,L...]. i ], where i = 1, 2, ..., n, and each daily load curve is denoted as L. i =[l i1 ,l i2 ,...,l i24 ]', l it This represents the load value in the t-th hour of the i-th day;

[0013] Meteorological Dataset D m D represents the factors affecting each load point. m =[D 1m D 2m ,...,D im ], where i = 1, 2, ..., n, D im D represents the sequence of m influencing factors on day i. i,m =[d 1,m ,d 2,m ,...,d 24,m ], d i,t,m This represents the value of influencing factor m at hour t on day i.

[0014] Step 2: Standardize the original data by normalizing it according to the following formula:

[0015]

[0016] Where X' represents the normalization result, X is the original data, Xmin represents the minimum value in the sequence, and X represents the maximum value in the sequence.

[0017] Step 3: Perform daily load characteristic classification processing, including calculating feature values, clustering based on daily load characteristics, and obtaining the clustering result V, specifically including the following steps:

[0018] Calculate eigenvalues:

[0019] Step 1: For each daily load sample L i Arranged in descending order, forming the load duration curve S i =[s i1, s i2 ,...,s i24 ];

[0020] Step 2: Let Lb i This represents the absolutely stable load value for each load curve under baseline conditions:

[0021]

[0022] Step 3: Calculate the characteristic value of each daily load curve, as shown below:

[0023] Calculate feature 1, for each S i Maximum load: f i1 =max(S) i )

[0024] Calculate feature 2, for each S i Load factor: f i2 =Lb i / f i1

[0025] Calculate feature 3, for each S i Hours above the baseline:

[0026] Calculate feature 4, for each S i Duration exceeding 50% of peak load:

[0027] Calculate feature 5, for each S i The duration exceeding 75% of the maximum load:

[0028] Calculate feature 6, for each S i The duration exceeding 95% of the maximum load:

[0029] Calculate feature 7, each S i The sum of the differences between all load values ​​greater than or equal to the baseline and the baseline:

[0030] Calculate feature 8, each S i The sum of the differences between all load values ​​less than or equal to the baseline and the baseline:

[0031] Calculate feature 9, each S i The rate of decline for the first n points (i.e., the highest load segment): f i9 =(s i1 -s iq ) / q;

[0032] Clustering based on daily load characteristics:

[0033] Step4: clustering the feature vectors F of each daily load curve i = [f i1 ,f i2 ,...,f i9 ] using the k-means++ method, and selecting the optimal clustering number K based on the silhouette coefficient, which is used to measure the clustering effect, and the silhouette coefficient of each sample point is calculated as follows:

[0034]

[0035] In the above formula, a(i) is the average distance from sample point i to other points in the same cluster, and b(i) is the average distance from sample point i to all points in a cluster that does not contain it; the silhouette coefficient of the sample is the average of the silhouette coefficients of all sample points After obtaining the silhouette coefficient of the sample, the adaptive K value selection is performed, and according to the adjusted K value, the K-means++ clustering is re-executed, and the clustering result is updated, and the iteration is performed multiple times until 0.5≤s(i)≤0.75, then the iteration is stopped;

[0036] Obtain the clustering result V:

[0037] The optimal result of K value is obtained by the above steps, and the optimal clustering result of K clusters is obtained, and all daily curves L i are classified, and labels type={1,2,...,K} are given respectively, and each i has a corresponding label V i ∈{1,2,...,K}, thereby forming the clustering variable V.

[0038] The step 4: causal feature discovery based on PC algorithm, including reconstructing the data set, causal feature discovery based on PC algorithm, and obtaining the causal feature set P t , specifically including the following steps:

[0039] Reconstruct the data set:

[0040] Step 1: for each daily load curve L i in the sample period, concatenate by date and time, denoted as L=[l 24(i-1)+t ], where t=1,2,...,24, i=1,2,...,n, l 24(i-1)+t denotes the load value at t time of the i-th day, thereby covering all sample quantities; concatenate the influence factor sequence D i,m corresponding to each load value in the same way, denoted as D m =[d 1,m ,d 2,m ,...,d (24(i-1)+t),m], where t = 1, 2,..., 24, i = 1, 2,..., n, d (24(i-1)+t),m denotes the value of the influence factor m at the i-th day and t-th hour;

[0041] Step2: Index the load sequence L by hour to obtain 24 reconstructed sub-sequences L t = [l 24*0+t , l 24*1+t ,..., l 24(n-1)+t ], L t denotes the load sub-sequence of the t-th hour in all days; In the same way, D m is reconstructed to obtain the influence factor sub-sequence D t,m = [d (24*0+t),m , d (24*1+t),m ,..., d (24*(n-1)+t),m ], D t,m denotes the sub-sequence of the t-th hour of the influence factor m in all days;

[0042] Step3: Process each D t,m to be approximately independent and identically distributed, and learn the approximate minimum weight W based on the following formula; D t,m multiplied by W can obtain D t,W which is approximately independent and identically distributed;

[0043] D t,w = W * D t,m

[0044]

[0045] where W ∈ R m×1 is the sample weight, ∑ w = diag(W1, W2,..., W m ), D t,-m denotes the variable except m in t-th hour, and when W reaches the minimum value, it indicates that the variables D t,m are independent of each other;

[0046] Causal feature discovery based on PC algorithm:

[0047] Step4: Take all D t,W and L t as ordinary variables, and use the local causal structure learning algorithm (HITON-PC) to learn the causal relationship PC t among all variables except L Lt ;

[0048] Step5: Screen the PC t,W containing only D LtThe feature set, the Fisher's z value is calculated based on the following formula as L t The current PC Lt The dependency value dep of the feature set, and find the minimum dependency value mindep; in the following formula, let The sample condition of given Z D t,m And L t The sample partial correlation coefficient;

[0049]

[0050] Step 6: select the corresponding L t There is a dependency relationship, and its dependency value dep is greater than or equal to the minimum dependency value mindep, add it to PCcan, thereby obtaining the best candidate feature set PCcan of each L t ;

[0051] Step 7: based on the best candidate feature set PCcan, for each candidate feature PC in it, conditional independence test is carried out according to Fisher's z value test, if there is a feature independent of L t , it is deleted from PCcan, and the final selected feature PCsel of each L t is obtained, that is, a completed partial directed acyclic graph (CPDAG),

[0052] Obtain the causal feature set P t：

[0053] Step 8: keep all the impact factors of the load node L t Directly point to, constitute the causal feature set P t .

[0054] Said step 5: input the PC-TCN prediction model based on daily load characteristic classification, and the specific method is that:

[0055] The load value at time t is predicted, based on the daily load characteristic classification in step 3, the classification label V is obtained, and based on the causal feature set P t Screened by the method in step 4, input the variable V, time point variable t, and feature set P t Into the time convolution network model, combine the time series convolution of TCN with the self-attention mechanism of Transformer, model the global long-term dependence while capturing local dependence;

[0056] Before the input layer of the TCN model, use the attention layer to weight each feature, and adaptively adjust the importance of the feature according to the load characteristics and clustering label V of different time steps; the feature weighting calculation formula is:

[0057] X't = X t * a t

[0058] wherein X' t is the weighted feature, X t is the original input feature, a t is the dynamic weight generated by the attention mechanism; the clustering label V is dynamically embedded into the feature input, and after being processed by the embedding layer, it is input into the model together with the time variable t and the feature set P t The load data at the tthhour is measured.

[0059] After the technology proposed in the present application is adopted, the short-term power load prediction method based on daily load characteristic clustering according to the embodiment of the present application has the following beneficial effects: the dimensionality of data is reduced by measuring the daily load characteristics, and the original date features are replaced by the obtained clustering results; at the same time, based on the causal discovery algorithm, the impact index set is screened, and the extraction of causal features is enhanced, so as to improve the operation efficiency and accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 A flowchart of the short-term power load prediction method based on daily load characteristic clustering according to the embodiment of the present application is shown. DETAILED DESCRIPTION

[0061] The various preferred embodiments of the present application will be described below with reference to the accompanying drawings. The following description with reference to the drawings is provided to help understand the example embodiments of the present application defined by the claims and their equivalents. It includes various specific details, but they can only be regarded as exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. Moreover, in order to make the description more concise and brief, detailed descriptions of functions and structures well known in the art will be omitted.

[0062] As Figure 1 shown, the short-term power load prediction method based on daily load characteristic clustering according to the embodiment of the present application includes the following steps:

[0063] Step 1: Obtain the original power load data, including:

[0064] The daily 24-point power load data set in a certain time period in the region is denoted as L = [L1, L2,..., Ln], where i = 1, 2,..., n, and each daily load curve is denoted as Li = [l1, l2,..., ln], where i = 1, 2,..., n. i i i1 i2 i24 it ​​​​​This represents the load value in the t-th hour of the i-th day.

[0065] Meteorological Dataset D m D represents the factors affecting each load point. m =[D 1m D 2m ,...,D im ], where i = 1, 2, ..., n, D im D represents the sequence of m influencing factors on day i. i,m =[d 1,m ,d 2,m ,...,d 24,m ], d i,t,m This represents the value of influencing factor m at hour t on day i.

[0066] The influencing factors include meteorological indicators such as point-in-time temperature, cumulative number of high-temperature days, cumulative number of low-temperature days, and humidity; a priori meteorological indicators such as load, temperature, cumulative number of high-temperature days, cumulative number of low-temperature days, and humidity in the h hours (h = 1 to 23 hours) before the point-in-time; and a priori meteorological indicators such as load, temperature, cumulative number of high-temperature days, cumulative number of low-temperature days, and humidity in the d days (d = 1 to 7 days) before the point-in-time, totaling 154 influencing characteristic indicators.

[0067] Step 2: Standardize the raw data

[0068] To avoid differences in units between different data, the original data is normalized:

[0069]

[0070] Where X' represents the normalization result, X is the original data, Xmin represents the minimum value in the sequence, and X represents the maximum value in the sequence.

[0071] The power load data is collected by sensors installed on sensor equipment to measure indicators such as voltage, current, power, and energy. The collected data is then transmitted to a monitoring terminal or cloud platform data center via a communication module. After receiving the power data, the data center saves it sequentially according to time series, thus forming a power data database. When forecasting power load, historical power data from the power data database is selected for use in power load forecasting.

[0072] The meteorological data is collected by sensors measuring meteorological elements such as temperature, humidity, air pressure, wind speed, wind direction, and precipitation, and then transmitted to the meteorological data acquisition system via wired or wireless means. When forecasting power load, historical meteorological data from a meteorological database is selected for use in power load forecasting.

[0073] Step 3: Perform daily load characteristic classification processing

[0074] Due to the difference of load curve at the same time point in different days, the representation ability of time index is affected to some extent, and characteristic redundancy is caused. Based on this, the application proposes a dimension reduction improved clustering algorithm based on curve characteristic value, which reduces the dimension of original data under the premise of combining the daily load characteristics, so as to improve the operation efficiency, realize the classification of the characteristics of all load samples in space, and then replace the corresponding date index into the subsequent model calculation, solve the problem of date characteristic redundancy, and the specific steps are as follows:

[0075] Calculate the characteristic value:

[0076] Step 1: For each daily load sample L i , arrange in descending order to form the load duration curve S i =[s i1 ,s i2 ,...,s i24 , at this time, the interference of time factor on the distribution of original sequence is eliminated, and the intensity distribution characteristics are focused.

[0077] Step 2: Let Lb i represent the absolute stable load value of each load curve under the reference condition:

[0078]

[0079] Step 3: Calculate the characteristic value of each daily load curve, as follows:

[0080] Calculate characteristic 1, the highest load of each S i : f i1 =max(S i )

[0081] Calculate characteristic 2, the load rate of each S i : f i2 =Lb i / f i1

[0082] Calculate characteristic 3, the number of hours higher than the reference line of each S i :

[0083] Calculate characteristic 4, the duration of each S i that exceeds 50% of the highest load:

[0084] Calculate characteristic 5, the duration of each S i that exceeds 75% of the highest load:

[0085] Calculate characteristic 6, the duration of each Si The duration exceeding 95% of the maximum load:

[0086] Calculate feature 7, each S i The sum of the differences between all load values ​​greater than or equal to the baseline and the baseline:

[0087] Calculate feature 8, each S i The sum of the differences between all load values ​​less than or equal to the baseline and the baseline:

[0088] Calculate feature 9, each S i The rate of decline for the first n points (i.e., the highest load segment): f i9 =(s i1 -s iq ) / q, where the value of q needs to be given based on industry experience.

[0089] Clustering based on daily load characteristics:

[0090] Step 4: Analyze the eigenvector F of each daily load curve. i =[f i1 ,f i2 ,...,f i9 Clustering was performed using the k-means++ method, and the optimal number of clusters K was selected based on the silhouette coefficient. The silhouette coefficient is used to measure the quality of clustering, and the specific formula for calculating the silhouette coefficient of each sample point is as follows:

[0091]

[0092] In the above formula, a(i) is the average distance from sample point i to other points within the same cluster, and b(i) is the average distance from sample point i to all points within a nearest cluster that does not contain it. This formula combines the clustering cohesion a(i) and separation b(i) to evaluate the clustering effect. This value ranges from -1 to 1; a larger value indicates closer proximity between samples within a cluster and greater distance between samples between clusters, indicating a better clustering effect. The silhouette coefficient of the sample is the mean of the silhouette coefficients of all sample points. After obtaining the silhouette coefficients of the samples, an adaptive K-value selection is performed. The adaptive K-value optimization steps are as follows:

[0093] Preliminary clustering: Set an initial K value of K0. Evaluate sample silhouette coefficients: Calculate sample silhouette coefficients based on the initial K value of K0. Avoid over-clustering or under-clustering by iterating the K value multiple times: If the sample silhouette coefficients reach a high value, such as s(i) > 0.75, consider reducing the K value. new= K0-1, to avoid too fine clustering, under-clustering: if the sample profile coefficient is low, such as s(i) < 0.5, the K value needs to be increased, K new = K0+1, to avoid different features of the sample being classified into the same cluster, if the sample profile coefficient is 0.5 < s(i) < 0.75, the current K value K0 is adopted. According to the adjusted K value, the K-means++ clustering is re-executed, and the clustering result is updated, and the iteration is performed for multiple times until 0.5 < s(i) < 0.75, and then the iteration is stopped;

[0094] Obtain the clustering result V:

[0095] The K value optimal result is obtained through the above steps, so that the optimal clustering result of K clusters is obtained, and the whole daily curve L i is classified, and a label type = {1, 2,..., K} is given, and for each i, there is a corresponding label V i ∈ {1, 2,..., K}, thereby forming a clustering variable V, which will be input into a prediction model subsequently.

[0096] Step 4: Causal feature discovery based on PC algorithm

[0097] Since the influence degree of each factor on the load value is significantly different at different time periods, the currently commonly used feature screening method generally directly measures the overall correlation between the feature sequence and the load sequence, and ignores the dynamic change of the causal strength of the load and the multiple factors at the hour granularity. Therefore, the application proposes a feature screening mechanism based on a causal discovery algorithm, which is used for the original influence factor D m The candidate feature set is constructed, and the local causal structure learning algorithm is introduced to solve the problem of insufficient processing of pseudo correlation.

[0098] The specific steps are as follows:

[0099] Reconstruct the data set:

[0100] Step 1: For each daily load curve L i in the sample period, date and time are spliced, denoted as L = [l1, l2,..., l 24(i-1)+t ], wherein t = 1, 2,..., 24, i = 1, 2,..., n, and l 24(i-1)+t represents the load value at t time of the i-th day, thereby covering the whole sample amount. The influence factor sequence D i,m corresponding to each load value is spliced in the same way, denoted as D m = [d 1,m , d 2,m ,..., d (24(i-1)+t),m ], wherein t = 1, 2,..., 24, i = 1, 2,..., n, and d (24(i-1)+t),mdenotes the value of the influence factor m at the i-th day and t-th time.

[0101] Step2: Index the load sequence L by hour, and get 24 reconstructed sub-sequences L t 24*0+t 24*1+t 24(n-1)+t t denotes the load sub-sequence at the t-th hour of all days. In the same way, reconstruct D m to get the influence factor sub-sequence D t,m (24*0+t),m (24*1+t),m (24*(n-1)+t),m t,m denotes the influence factor m sub-sequence at the t-th hour of all days.

[0102] Step3: Process each D t,m to be approximately independent and identically distributed, and learn the approximate minimum weight W based on the following formula. Multiply the original data D t,m by W to get D t,W which is approximately independent and identically distributed.

[0103] D t,w = W * D t,m

[0104]

[0105] where W ∈ R m×1 is the sample weight, Σ w = diag(W1,W2,...,W m ), D t,-m denotes the variable at t-th time except m, and when W reaches the minimum value, it indicates that the variables D t,m are independent of each other.

[0106] Causal feature discovery based on PC algorithm:

[0107] Step4: Take all D t,W and L t as ordinary variables, and use the local causal structure learning algorithm (HITON-PC) to learn the causal relationship PC t among all variables except L Lt .

[0108] Step5: Screen the PC t,W feature set containing only D Lt variables, and calculate the Fisher'z value based on the following formula to be used as L t and the current PC​​​​​​​​Lt The dependency value dep of the feature set is found, and the minimum dependency value mindep is found. In the following formula, let D t,m and the sample partial correlation coefficient of L t .

[0109]

[0110] Step 6: Select the features that have a dependency relationship with the corresponding L t , and the dependency value dep is greater than or equal to the minimum dependency value mindep, and add them to PCcan, thereby obtaining the optimal candidate feature set PCcan for each L t .

[0111] Step 7: Based on the optimal candidate feature set PCcan, for each candidate feature PC in it, a conditional independence test is performed according to the Fisher'z value test, and if there is a feature independent of L t , it is deleted from PCcan, and the final filtered feature PCsel for each L t is obtained, that is, a completed partial directed acyclic graph (CPDAG),

[0112] obtain the causal feature set P t：

[0113] Step 8: Keep all the impact factors that directly point to the load node L t , and construct the causal feature set P t .

[0114] Step 5: Input the PC-TCN prediction model based on daily load characteristic classification

[0115] Predict the load value at time t, based on the daily load characteristic classification in step 3, obtain the classification label V, and based on the causal feature set P t filtered by the method in step 4, input the variables V, time point variable t, and feature set P t into the Temporal Convolutional Network (TCN) model, but the TCN has weak ability to capture long-term dependencies. To overcome this shortcoming, the TCN and the Transformer model need to be combined. By combining the time convolution of the TCN and the self-attention mechanism of the Transformer, local dependencies can be captured while modeling global long-term dependencies. The model architecture is as follows: TCN part: maintain the dilated convolution structure to capture local temporal dependencies, Transformer part: add a self-attention layer after the output of the TCN to capture long-term temporal dependencies.

[0116] At the same time, considering the weight difference of different features at different time periods, an attention layer is used to weight each feature before the input layer of the TCN model, and the importance of the feature is adaptively adjusted according to the load characteristics and the clustering label V at different time steps. The feature weighting calculation formula is:

[0117] X′ t =X t ×α t

[0118] Wherein, X′ t is the weighted feature, X t is the original input feature, and a t is the dynamic weight generated by the attention mechanism. The clustering label V is dynamically embedded into the feature input, and after processing by the embedding layer, it is input into the model together with the time point variable t and the feature set P t In this way, the weight difference of the feature at different time points is effectively considered, avoiding the limitation of static weight setting, and improving the robustness and timeliness of the model.

[0119] Finally, the improved TCN model can predict the measurable load data at the tth hour. The output prediction result is output.

[0120] The above describes the present application in detail, and the principle and implementation mode of the present application are described by applying specific examples; the above example is only used to help understand the method and core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed, and the above description should not be understood as the limitation of the present application.

[0121] Through the description of the above implementation mode, those skilled in the art can clearly understand the implementation of the present application. Of course, the above listed cases are only examples, and the present application is not limited to this. Those skilled in the art should understand that other deformations or simplifications of the technical solutions of the present application can be appropriately applied to the present application, and should be included in the scope of the present application.

Claims

1. A short-term power load forecasting method based on daily load characteristics clustering, characterized by, The method comprises the following steps: Step 1: obtaining power load original data; Step 2: standardizing the original data; Step 3: performing daily load characteristic classification processing; Step 4: discovering causal characteristics based on a PC algorithm; Step 5: inputting a PC-TCN prediction model based on daily load characteristic classification; Step 6: measuring and outputting a prediction result.

2. The short-term power load forecasting method based on daily load characteristics clustering according to claim 1, characterized in that, The power load original data is obtained by: The daily 24-hour electricity load dataset for a certain time period within a region is denoted as L=[L1,L2,...,L...]. i ], where i = 1, 2, ..., n, and each daily load curve is denoted as L. i =[l i1 ,l i2 ,...,l i24 ]', l it This represents the load value in the t-th hour of the i-th day; Weather dataset D m D m = [D 1m ,D 2m ,...,D im ], where i = 1, 2,..., n, D im represents a sequence of m influencing factors for the i-th day, D i,m = [d 1,m ,d 2,m ,...,d 24,m ], d i,t,m represents the m influencing factor values for the t-th hour of the i-th day. 3.The short-term power load forecasting method based on daily load characteristics clustering of claim 1, wherein, The original data is standardized by normalization according to the following formula: Wherein, X' represents the normalized result, X is the original data, Xmin represents the minimum value in the sequence, and X represents the maximum value in the sequence. 4.The short-term power load forecasting method based on daily load characteristics clustering of claim 1, wherein, The daily load characteristic classification processing comprises calculating characteristic values, clustering according to daily load characteristics, and obtaining a clustering result V, and specifically comprises the following steps: Calculating characteristic values: Step 1: For each daily load sample L i In descending order, form the load duration curve S i = [s i1 ,s i2 ,...,s i24 ] ; Step 2: Let Lb i denote the absolute stationary load value of each load curve in the base case: Step 3: calculating the characteristic values of each daily load curve, as follows: Compute feature 1, each S i Maximum load of: f i1 = max(S i ) Compute feature 2, each S i Load rate of: f i2 = Lb i / f i1 Compute feature 3, each S i hours above baseline: Compute feature 4, each S i Duration of time over 50% of the highest load: Calculate feature 5, each S i Duration of time over 75% of the highest load: Compute feature 6, each S i Duration of time over 95% of the highest load: Compute feature 7, each S i Sum of the difference between all load values greater than or equal to the baseline and the baseline: Compute feature 8, each S i Sum of the difference between all load values less than or equal to the baseline and the baseline: Compute feature 9, each S i Slope of descent of the first n points (i.e. the highest load segment): f i9 = (s i1 -s iq ) / q; Clustering according to daily load characteristics: Step 4: Analyze the eigenvector F of each daily load curve. i =[f i1 ,f i2 ,...,f i9 Clustering was performed using the k-means++ method, and the optimal number of clusters K was selected based on the silhouette coefficient. The silhouette coefficient is used to measure the quality of clustering, and the specific formula for calculating the silhouette coefficient of each sample point is as follows: In the above formula, a(i) is the average distance from sample point i to other points in the same cluster, and b(i) is the average distance from sample point i to all points in a nearest cluster not containing it; the profile coefficient of the sample is the average of the profile coefficients of all sample points After obtaining the profile coefficient of the sample, adaptive K value selection is performed, K-means++ clustering is re-executed according to the adjusted K value, and the clustering result is updated, and the iteration is performed multiple times, until 0.5≤s(i)≤0.75, and then the iteration is stopped. Obtaining a clustering result V: The optimal result of K value is obtained from the above steps, so as to obtain the optimal clustering result of K clusters, and the whole daily curve L i Classification is performed, and labels type = {1, 2,..., K} are respectively given, and each i has a corresponding label V i ∈{1, 2,..., K}, thereby forming a clustering variable V.

5. The short-term power load forecasting method based on daily load characteristics clustering according to claim 1, characterized in that, The PC algorithm-based causal feature discovery includes reconstructing a data set, PC algorithm-based causal feature discovery, and obtaining a causal feature set P t , and specifically includes the following steps: Reconstructing a data set: Step 1: For each daily load curve L in the sample period i , concatenate by date-time, denoted as L = [l 24(i-1)+t , l 24(i-1)+t ,..., l i,m ], where t = 1, 2,..., 24, i = 1, 2,..., n, l m denotes the load value at time t on the i-th day, thus covering the total sample size; concatenate the impact factor sequence D corresponding to each load value in the same way, denoted as D = [d 1,m , d 2,m ,..., d (24(i-1)+t),m ], where t = 1, 2,..., 24, i = 1, 2,..., n, d (24(i-1)+t),m denotes the value of the impact factor m at time t on the i-th day; Step 2: Index the load sequence L by hour to obtain 24 reconstructed sub-sequences L t = [l 24*0+t , l 24 ,..., l 24(n-1)+t ], L t denotes the load sub-sequence of the t-th hour in all days; in the same way, D m is reconstructed to obtain the impact factor sub-sequence D t,m = [d (24*0+t),m , d (24*1+t),m ,..., d (24*(n-1)+t),m ], D t,m denotes the impact factor m sub-sequence of the t-th hour in all days; Step3: For each D t,m approximate independent and identically distributed, learn the approximate minimum weight W based on the following formula; from the original data D t,m multiply by W to get D t,W is approximately independent and identically distributed; D t,w = W * D t,m where W ∈ R m×1 is the sample weight, Σ w = diag(W1, W2,..., W m ), D t,-m represents the variables other than m at time t, when W reaches the minimum value, it indicates that the variables D t,m are independent of each other; Discovering causal characteristics based on a PC algorithm: Step4: All D t,W with L t are treated as ordinary variables, and the local causal structure learning algorithm (HITON-PC) is used to learn causal relationships among all variables except L t PC Lt ; Step5: Select only D t,W PC of variable Lt Feature set, Fisher's z value is calculated based on the following formula as L t With the current PC Lt Dependency value dep of feature set, and find the minimum dependency value mindep; in the following formula, let D t,m and L t Sample partial correlation coefficient of Step6: select the L t The features that have dependency relationship and whose dependency value dep is greater than or equal to the minimum dependency value mindep are added into the PCcan, thus obtaining the optimal candidate feature set PCcan of each L t . Step 7: Based on the best candidate feature set PCcan, for each candidate feature PC in it, a conditional independence test is performed according to Fisher's z-value test, and if there exists a feature that is independent of L t , then it is removed from PCcan, resulting in each L t The final selected features PCsel, i.e., a completed partially directed acyclic graph (CPDAG), obtaining a causal feature set P t： Step 8: Keep all direct pointing to load nodes L t The influencing factors of the above are the causal feature set P t .

6. The short-term power load forecasting method based on daily load characteristics clustering according to claim 1, characterized in that, The PC-TCN prediction model based on daily load characteristic classification is inputted by: The load value at time t is predicted, and a classification label V is obtained based on the daily load characteristic classification in step 3, and a causal feature set P is obtained based on the method screening in step 4 t The variable V, the time point variable t, and the feature set P t In the input time convolution network model, the time series convolution of TCN is combined with the self-attention mechanism of Transformer, so as to model the global long-term dependence while capturing local dependence. Before the input layer of the TCN model, an attention layer is used to weight each feature, and the importance of the feature is adaptively adjusted according to the load characteristics and the clustering label V of different time steps; the feature weighting calculation formula is as follows: X' t = X t x a t wherein X′ t is the weighted feature, X t is the original input feature, a t is the dynamic weight generated by the attention mechanism; the clustering label V is dynamically embedded into the feature input, processed by the embedding layer, and input into the model together with the time variable t and the feature set P t The load data at the tth hour is measured.

Images