Short-term load prediction method based on time sequence importance analysis and feature extraction

Abstract

The invention discloses a short-term load prediction method based on time sequence importance analysis and feature extraction, and the method comprises the steps: carrying out the clustering analysis of a power load data set with a day as a unit, and taking each clustering center as a typical representative daily load curve based on a clustering result; extracting morphological characteristics of the time sequence data corresponding to each type of typical representative daily load curve through an extreme point importance extraction algorithm to obtain an extreme point importance characteristic sequence of each type of typical representative daily load curve; selecting external factor input features based on historical meteorological data and date data in the power data set, and reconstructing the external factor input features and extreme point importance feature sequences of each type of typical representative daily load curves into a new feature set; and inputting the new feature set of each type of load into the CNN-LSTM hybrid model for prediction to obtain the prediction load of each type of prediction day. The overall load prediction precision is improved by improving the prediction precision of the prediction model for the extremal region of each typical representative day.

Description

Technical Field

[0001] The present invention relates to the technical field of electric load forecasting, and in particular to a short-term load forecasting method. Background Art

[0002] Load forecasting refers to the technology of predicting and estimating the load demand in a power system. Based on the historical data of the electric load, combined with basic data such as historical weather data, precipitation and other weather impacts, the prosperity and depression of the economy, and the social status quo, the internal relationship between the load and other relevant factors is explored to achieve the purpose of accurately forecasting the load. At present, short-term load forecasting methods are mainly divided into two categories: classical forecasting methods and machine learning methods. In the machine learning method, in order to improve the generalization performance of the deep learning model, appropriate feature variables need to be selected to enrich the input of the forecasting model.

[0003] According to the fact that electric load data and various factors affecting the load have time-related characteristics, time series data mining algorithms are widely used in load forecasting to exploratively obtain various valuable patterns or rules from the time series. Mining and analyzing the characteristics of time series data is one of the methods to improve the forecasting accuracy. Existing short-term load forecasting research and extreme value forecasting on time series mainly focus on four aspects: (1) Using a decomposition algorithm for load data and then forecasting each component. (2) Extracting specific extreme value data sets for corresponding training to achieve high-precision load forecasting. (3) Improving the forecasting model, adopting a model more suitable for time series or using a combined forecasting model. The combined forecasting model combines multiple models and methods, and better meets the actual needs of short-term electric load forecasting by combining the characteristics and advantages of different models. (4) By improving the overall forecasting accuracy, improving the accuracy improvement in the extreme value region. Although most literatures recognize the impact of time series on the load, the consideration of peak-valley extreme points is still insufficient. These methods simply extract the peak-valley data characteristics and do not conduct targeted research on the extreme points by combining the curve characteristics. The improvement of the extreme point accuracy is usually based on the improvement of the overall accuracy of extracting historical peak data or the forecasting result.

[0004] For example, the invention with the application number 202411123247.2 discloses a short-term electric load forecasting method based on modal aggregation and optimal integration, including: obtaining the electric load data in the power system; preliminarily decomposing the load sequence based on ICEEMDAN; quantifying the complexity of the subsequences by permutation entropy, and reconstructing the subsequences with high entropy value similarity through K-medoids clustering to obtain the aggregated modal components; further denoising the high-frequency noise components in the aggregated modal components by using the Savitzky-Golay filter; using BiGRU as a predictor to train and predict the aggregated modal components respectively; constructing a weight optimization model with the prediction accuracy and stability as the goals; solving the optimization model by NSDBO to obtain the Pareto front; using TOPSIS as a decision-making method to calculate the relative closeness of each Pareto optimal solution, and determining the best weight scheme; and weighted superposing the aggregated modal prediction sequences based on the decision-making scheme to obtain the final short-term electric load forecasting result. This patent improves the overall performance of short-term load forecasting through the decomposition and combination model, but its technical path still continues the limitations of traditional methods - oriented to global accuracy and lacking targeted modeling of the dynamic characteristics of extreme points and external incentives. Summary of the Invention

[0005] Aiming at the technical problem that few existing load forecasting methods extract the technical features that can represent the change characteristics of load data by analyzing various characteristics of the load data itself, the present invention starts from mining the potential morphological features of time series data, and combines the advantages of the combined forecasting model based on the deep learning method to propose a short-term load forecasting method based on time series importance analysis and feature extraction. By performing multiple loops on the extreme value interval region through the proposed extreme point importance extraction algorithm, the importance of each extreme point is obtained to different degrees, effectively converting the abstract and unavailable curve features into data vectors, improving the load forecasting accuracy in the extreme point region, and thus realizing the improvement of the load forecasting accuracy.

[0006] In order to achieve the above object, the technical solution of the present invention is realized as follows:

[0007] A short-term load forecasting method based on time series importance analysis and feature extraction, including the following steps:

[0008] S1. Perform clustering analysis on the electric load data set containing historical load data, historical meteorological data, and date data on a daily basis, and based on the clustering results, use each clustering center as the typical representative daily load curve;

[0009] S2. Extract the morphological features of the time series data corresponding to the typical representative daily load curve of each category through the extreme point importance extraction algorithm to obtain the extreme point importance feature sequence of the typical representative daily load curve of each category;

[0010] S3. Based on the historical meteorological data and date data in the power dataset, select the external factor input features, reconstruct the external factor input features and the extreme point importance feature sequences of the typical representative daily load curves of each category into new feature sets, and input the new feature sets of each category of load into the CNN-LSTM hybrid model for prediction to obtain the predicted loads of each type of prediction day.

[0011] Preferably, the extreme point importance extraction algorithm consists of two parts: an extreme point judgment algorithm for judging whether the elements in the sequence are extreme points and marking them; an importance marking algorithm for marking the importance of each element in the sequence by level and outputting the extreme point importance feature sequence.

[0012] Preferably, the method for obtaining the extreme point importance feature sequence of the typical representative daily load curve of each category is as follows: During the process of traversing the typical representative daily load curve using the importance marking algorithm according to the marks, in each loop, first use the extreme point judgment algorithm to screen and mark the extreme points of the typical representative daily load curve, and then judge the importance level according to the number of marks to extract the extreme point importance feature sequence of the typical representative daily load curve.

[0013] Preferably, the specific method for first using the extreme point judgment algorithm to screen and mark the extreme points of the typical representative daily load curve is as follows:

[0014] Suppose there is a sequence S = {s1, s2, s3, s4, …, s n}, the sequence length is n, the m-th element of the sequence S is s m , v represents the current calculation neighborhood; if the element s m is larger than the v elements in front of the element s m and the v elements behind the element s m , then the mark of the element s m is 1; if the element s m is smaller than the v elements in front of the element s m and the v elements behind the element s m , then the mark of the element s m is -1, otherwise the mark of the element s m is taken as 0, and the formula is as follows:

[0015]

[0016] Among them, Flag is the extreme value mark, and k is the range of the extreme value calculation neighborhood.

[0017] Preferably, the method for extracting the extreme point importance feature sequence of the typical representative daily load curve is as follows:

[0018] For the sequence S = {s1, s2, s3, s4, …, s n}, within the range of the first k elements before and after the neighborhood, perform the importance flag algorithm calculation: First, initialize the extreme point importance feature sequence, and then traverse the sequence S through two nested loops; the outer loop ranges from 1 to k, and the inner loop traverses each element of the sequence S; in each inner loop, judge the degree of association between each point and the extreme point according to the mark of the extreme point judgment algorithm and the value of the extreme point importance feature sequence. If the condition is satisfied where a represents the outer loop, b represents the inner loop, TAG(b) represents the value of the TAG sequence in the current loop, 1 ≤ a ≤ k, 1 ≤ b ≤ 24, then accordingly update the value of the extreme point importance feature sequence; after completing the two nested loops, obtain the updated extreme point importance feature sequence; the extreme point importance feature sequence TAG = {z1, z2, z3, …, z n}, where z m represents the importance degree of the m-th element in the S sequence.

[0019] Preferably, use the K-Medoids algorithm for clustering analysis, select the cosine similarity as the metric index, and use the silhouette coefficient as the evaluation index, and select the K value with the highest overall silhouette coefficient as the optimal number of clusters.

[0020] Preferably, the calculation method of the cosine similarity is as follows:

[0021] Let x i = [x i1 , x i2 , …, x in and x j = [x j1 , x j2 , …, x jn be the load data sequences of the i-th day and the j-th day of the load in a certain area respectively. Then the calculation formula of the cosine similarity of the load data sequences of the i-th day and the j-th day is as follows:

[0022]

[0023] where S cos (x i , x j ) is the cosine similarity of the load data sequences of the i-th day and the j-th day, x iq is the q-th element in the sequence x i , x iq is the q-th element in the sequence x j , and n is the sequence length.

[0024] Preferably, the calculation method of the overall silhouette coefficient is as follows:

[0025] First, calculate the sequence x i and the sequence x j of the silhouette coefficient, and the calculation formula is as follows:

[0026]

[0027] In the formula, a(i) is the average cosine similarity of the sequence x i to other sequences x of the same category j ; b(i) is the average cosine similarity of the sequence x i to the sequences x in other categories j ;

[0028] Next, calculate the mean of the silhouette coefficients of all sequences to obtain the overall silhouette coefficient of the clustering result, and the formula is as follows:

[0029]

[0030] In the formula, the overall silhouette coefficient S C takes values in [-1, 1], and N is the number of all sequences.

[0031] Preferably, the specific structure of the CNN-LSTM hybrid model includes, connected in sequence: Convolution Layer I, Max Pooling Layer I, Convolution Layer II, Max Pooling Layer II, Convolution Layer III, Max Pooling Layer III, Flatten Layer, LSTM Layer I, Dropout Layer I, LSTM Layer II, Dropout Layer II, and Fully Connected Layer.

[0032] Preferably, the external factor input features include temperature, humidity, temperature-humidity index, and date type.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] The present invention provides an innovative method for load forecasting. The proposed extreme point importance extraction algorithm performs multiple loops on the extreme value interval region to obtain the importance of each extreme point to different degrees, thereby not only reflecting the overall trend characteristics but also highlighting the extreme characteristics. By combining the external factor features, the prediction model pays more attention to the extreme value region during the machine learning training process, thus improving the prediction accuracy.

[0035] The present invention captures the morphological features of the load curve, especially the interval features between extreme points, through the marking and calculation of the importance of extreme points. By fusing the morphological features of the load curve in the feature set, the prediction accuracy of the extreme value region of each typical representative day of the prediction model is improved, and thus the overall load prediction accuracy is improved. Description of the Drawings

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the accompanying drawings required in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0037] Figure 1 Schematic diagram of the extreme point importance extraction algorithm structure of the present invention.

[0038] Figure 2 Extreme points from October 8th to 12th in an embodiment of the present invention.

[0039] Figure 3 Flowchart of the importance marking algorithm of the present invention.

[0040] Figure 4 Schematic diagram of the structure of the CNN-LSTM hybrid model.

[0041] Figure 5 Overall schematic diagram of the present invention.

[0042] Figure 6 Clustering result in an embodiment of the present invention, where Figure 6 -(a) is the working day type, Figure 6 -(b) is the high temperature day type, Figure 6 -(c) is the weekend / holiday type.

[0043] Figure 7 Importance of each type in an embodiment of the present invention, where Figure 7 -(a) is the importance of the working day, Figure 7 -(b) is the importance of the high temperature day, Figure 7 -(c) is the importance of the weekend / holiday.

[0044] Figure 8 Prediction results of each type of day in an embodiment of the present invention, where Figure 8 -(a) is the prediction result from October 21st to 25th for the working day, Figure 8 -(b) is the prediction result from August 19th to 22nd for the high temperature day, Figure 8 -(c) is the prediction result from November 23rd to 24th for the weekend / holiday.

[0045] Figure 9 Error radar chart of each type of day in an embodiment of the present invention, where Figure 9 -(a) is the working day radar chart, Figure 9 -(b) is the high temperature day radar chart, Figure 9 -(c) is the holiday radar chart.

[0046] Figure 10 For the absolute daily errors of various types in an embodiment of the present invention, where Figure 10 -(a) is the absolute error on weekdays, Figure 10 -(b) is the absolute error on high-temperature days, Figure 10 -(c) is the absolute error on weekend days. Detailed implementation manners

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0048] To improve the prediction accuracy and the generalization ability of the prediction model, the present invention proposes a hybrid prediction model K-E-CNN-LSTM of convolutional neural networks and long short-term memory (CNN-LSTM) that extracts extreme value importance based on the EIIR (Enhanced Importance Index Recognize) algorithm. First, the K-Medoids algorithm is applied to cluster the daily load curves, then the EIIR algorithm is used to extract the numerical features of the extreme value points of each cluster center. Finally, the importance feature sequences of the extreme value points of each typical representative daily load curve and the corresponding historical load type data are reconstructed into a new feature set, and the new feature set is input into the CNN-LSTM hybrid model for prediction to obtain the prediction result.

[0049] A short-term load prediction method based on time series importance analysis and feature extraction, comprising the following steps:

[0050] S1. To reflect the characteristics of different scenarios of load data, a power load data set containing historical load data, historical meteorological data, and date data is clustered on a daily basis. Based on the clustering results, each cluster center is used as a typical representative daily load curve.

[0051] Further, the K-Medoids algorithm is used for clustering analysis, the cosine similarity is selected as the metric index, and the silhouette coefficient is used as the evaluation index. The K value with the highest overall silhouette coefficient is selected as the optimal number of clusters.

[0052] The key of the K-Medoid algorithm lies in the selection of cluster center points. The criterion for selecting cluster center points is that the sum of the distances from all other points in the current cluster to this center point is the smallest. Therefore, all points in the cluster need to be traversed. Since the Euclidean distance cannot accurately reflect the morphological characteristics of the load curve, when the amplitudes of the load curves are similar but there are morphological differences, the cosine similarity can better identify the morphological characteristics of the load curve. Therefore, the cosine similarity is selected as the metric for classifying daily load characteristics.

[0053] Let x i = [x i1 , x i2 , …, x in and x j = [x j1 , x j2 , …, x jn be the load data sequences of the i-th day and the j-th day of the load in a certain area respectively. Then the calculation formula for the cosine similarity of the load data sequences of the i-th day and the j-th day is as follows:

[0054]

[0055] Among them, the larger the cosine similarity S cos (x i , x j ) of the load data sequences of the i-th day and the j-th day, the more similar the daily load curve shapes represented by the sequences x i and the sequence x j are, and the greater the chance of belonging to the same class. When the cosine similarity S cos (x i , x j ) of the load data sequences of the i-th day and the j-th day is 1, the daily load curves represented by the sequences x i and the sequence x j completely coincide. x iq is the q-th element in the sequence x i , and x iq is the q-th element in the sequence x j , and n is the sequence length.

[0056] To determine the optimal number of clusters, it is necessary to conduct a clustering validity test on the clustering results. The conventional silhouette coefficient method is used as the evaluation index, and the calculation formula is as follows:

[0057]

[0058] In the formula, S c (i) is the silhouette coefficient of the sequences x i and the sequence x j , and a(i) is the distance from the sequence x i to other sequences xj Average cosine similarity; b(i) is the sequence x i to the sequence x in other classes j Average cosine similarity. The mean of the silhouette coefficients of all sequences is calculated to obtain the overall silhouette coefficient of the clustering result, and the formula is as follows:

[0059]

[0060] In the formula, the overall silhouette coefficient S C takes values in [-1, 1]. The closer the value of the overall silhouette coefficient S C is to 1, the better the clustering effect. N is the number of all sequences.

[0061] S2. The morphological features of the time series data corresponding to the typical representative daily load curves of each class are extracted through the extreme point importance extraction algorithm to obtain the extreme point importance feature sequences of the typical representative daily load curves of each class.

[0062] Furthermore, as Figure 1 shown, the extreme point importance extraction algorithm proposed by the present invention based on the data characteristics of time series consists of two parts: the extreme point judgment algorithm (JEP), which is used to judge whether the elements in the sequence are extreme points and mark them. The importance marking algorithm (EIIR) is used to mark the importance of each element in the sequence by rank and output the extreme point importance feature sequence.

[0063] Furthermore, in the process of traversing the typical representative daily load curves by using the importance marking algorithm according to the marks, in each loop, the extreme point judgment algorithm is first used to screen and mark the extreme points of the typical representative daily load curves, and then the importance level is judged according to the number of marks, and the extreme point importance feature sequences of the typical representative daily load curves are extracted to obtain the morphological features of the time series data corresponding to the typical representative daily load curves.

[0064] Specifically, the specific method of first using the extreme point judgment algorithm (JEP) to screen and mark the extreme points of the typical representative daily load curves is as follows:

[0065] Suppose there is a sequence S = {s1, s2, s3, s4,..., s n}, the length of the sequence is n, the m-th element of the sequence S is s m , and v represents the current calculation neighborhood. If the element s m is greater than the previous s m s m-v+1 , s m-v ,... and other v elements as well as the s m behind the element s m+v-1 , s m+v ... and other v elements, then the element sm is marked as 1; if element s m is smaller than s m before s m-v+1 , s m-v , … and the v elements such as s and s m after s m+v-1 , s m+v … and the v elements such as s are all smaller, then the mark of element s m is marked as -1, otherwise the mark of element s m is taken as 0, and the formula is as follows:

[0066]

[0067] Among them, Flag is the extreme value mark, and k is the calculation neighborhood range of the extreme value, which is used to determine the extreme value mark interval. For both endpoints or where the neighborhood is less than v, the insufficient neighborhood side is replaced by the endpoint, and the other side is judged normally.

[0068] Taking the daily load curves of 5 days from October 8th to October 12th, 2012 in the ISO-NE public power load dataset as an example. Each daily load curve contains many local extreme points. Assume that the current calculation neighborhood v = 6, and the extreme points marked by the extreme point judgment algorithm are as Figure 2 shown.

[0069] Specifically, the method for outputting the extreme point importance feature sequence through the importance flag algorithm (EIIR) is as follows:

[0070] For the sequence S = {s1, s2, s3, s4, …, s n}, the EIIR algorithm is calculated within the range of the k elements before and after the neighborhood. As Figure 3 shown, a in the figure is the outer loop, that is, k neighborhoods, and b is the inner loop, that is, 24 time points. First, initialize the extreme point importance feature sequence TAG, and then traverse the sequence S through the inner and outer loops. The outer loop goes from 1 to k, and the inner loop traverses each element of the sequence S. In each inner loop, according to the mark of the JEP algorithm and the value of the extreme point importance feature sequence, judge the association degree of each point and the extreme point. If the condition is met (where a represents the outer loop, b represents the inner loop, and TAG(b) represents the value of the TAG sequence in the current loop, 1 ≤ a ≤ k, 1 ≤ b ≤ 24), then update the value of the extreme point importance feature sequence accordingly. After completing the inner and outer loops, the updated extreme point importance feature sequence is obtained.

[0071] The extreme point importance extraction algorithm takes a sequence as input and applies the EIIR algorithm in a iterations to accumulate the extreme point importance of each point in the sequence. In each outer loop iteration, the neighborhood range increases by one unit (a increases from 1 to k), and then the extreme point flag Flag of each point in this neighborhood range is recalculated, and the extreme point importance feature sequence is updated. Simply put, the main idea of the extreme point importance extraction algorithm is that the extreme point importance of a point is determined by the frequency with which it is identified as an extreme point. For example, if element s3 is the maximum value in neighborhoods of sizes 1, 2, and 3 respectively, then after the EIIR algorithm, the corresponding TAG[3] will be added as 3. Similarly, if element s3 is the minimum value in neighborhoods of sizes 1, 2, and 3 respectively, the corresponding TAG[3] will be decreased to -3. If it is the maximum value in only one of the six neighborhood ranges, then finally TAG[3] will output 1. Thus, the importance of each extreme point is shown.

[0072] The extreme point importance feature sequence TAG = {z1, z2, z3, …, z n}, where z m represents the importance degree of the m-th element in the S sequence.

[0073] Select the load from October 21st to 25th, 2013 on weekdays, the load from August 19th to 22nd on high-temperature days, and the load from November 23rd to 24th on weekends / holidays in the ISO-NE public power load dataset as examples. Use different k values to extract features from the load curves of three typical representative days, namely weekdays, high-temperature days, and weekends / holidays, after K-Medoids clustering, and use CNN and LSTM models for prediction respectively. The prediction models are abbreviated as K-E-CNN model and K-E-LSTM model.

[0074] In order to fully reflect the extreme point characteristics of the load curve and verify the influence of different k values on each typical representative day, in the load prediction models of the K-E-CNN model and the K-E-LSTM model, the calculation neighborhood range k of the extreme value is taken as 3, 6, and 9 respectively, indicating that the EIIR calculation range is within k points before and after. The results are shown in Table 1.

[0075] Table 1 Prediction results for different k values

[0076]

[0077] From the prediction results in Table 1, it can be obtained that among the three types of days, when the parameter k is taken as 6, both the K-E-CNN model and the K-LSTM model reach a relatively high prediction accuracy. Therefore, in this embodiment, when using the K-E-CNN-LSTM hybrid prediction model for prediction, the calculation neighborhood range k of the extreme value is taken as 6.

[0078] S3. Based on the historical meteorological data and date data in the power data concentration, select the external factor input features, reconstruct the external factor input features and the extreme point importance feature sequences of the typical representative day load curves of each category into new feature sets respectively, and input the new feature sets of each type of load into the CNN-LSTM hybrid model for prediction to obtain the predicted loads of each type of prediction day.

[0079] In load forecasting, there are redundant feature inputs in the original load data. If inappropriate features are not selected as inputs, it will waste computing resources, reduce the efficiency of the forecasting model, and also affect the learning ability of the forecasting model network, learning unnecessary things. Therefore, feature screening is required.

[0080] Meteorological factors have a crucial impact on short-term load forecasting. Among them, the common influencing factors are temperature, humidity, or temperature-humidity index, weather type, etc. Loads change sharply in summer and winter due to changes in human living behaviors.

[0081] The date type is another important influencing factor in short-term load forecasting. The power loads on non-working days, namely Saturdays, Sundays, and holidays, are quite different from those on working days, showing a decreasing trend. And the high load caused by the continuous high temperature in summer also makes the load different from other periods. Therefore, the date type is listed as one of the factors affecting the load forecasting results.

[0082] For the convenience of quantitative calculation, in the feature of week type, the numbers 1 to 7 are used to represent Monday to Sunday. In the feature of holiday type, the number 1 is used to mark working days, and the number 2 is used to mark holidays and weekends. These are two independent and different feature sequences. Using the numbers 1 and 2 is to quantify the features, and the same markings do not affect training.

[0083] In summary, the external factor input features are shown in Table 2, and the new feature set is composed of the external factor input features and the extreme point importance feature sequences.

[0084] Table 2 External factor input features

[0085]

[0086] As Figure 4 and Figure 5 shown, the CNN-LSTM hybrid model adopted in the present invention is composed of 3 one-dimensional CNN layers and 2 LSTM layers. The specific structure includes, connected in sequence: Convolution Layer I, Max Pooling Layer I, Convolution Layer II, Max Pooling Layer II, Convolution Layer III, Max Pooling Layer III, Flatten Layer, LSTM Layer I, Dropout Layer I, LSTM Layer II, Dropout Layer II, and Fully Connected Layer.

[0087] The CNN-LSTM hybrid model is a hybrid neural network that combines the feature extraction ability of CNN and the long-term memory ability of LSTM for time series. The CNN is used to mine the features between data to form new feature vectors, and then these feature vectors are input into the LSTM for prediction. For the data input into the CNN-LSTM hybrid model, first, the convolutional layer of the CNN extracts local features, and the extracted feature vectors are passed to the max pooling layer for downsampling of the feature vectors and compression of the data volume. Then, the feature vectors processed by the convolutional layer and the max pooling layer are transformed into one-dimensional vectors through a flattening layer and input into the LSTM layer for prediction. Among them, the dropout layer after the LSTM layer is used to prevent the model from overfitting.

[0088] The parameter settings of the CNN-LSTM hybrid model are as follows: the filter size in Convolutional Layer Ⅰ is [3 1], and the number of filters is 64; the filter size in Convolutional Layer Ⅱ is [3 1], and the number of filters is 32; the filter size in Convolutional Layer Ⅲ is [3 1], and the number of filters is 16; the max pooling layer size is [2 1]; the number of network units in LSTM Layer Ⅰ is selected as 100, and the number of network units in LSTM Layer Ⅱ is selected as 50. The dropout probability of the dropout layer is all set to 0.2. The optimization algorithm is the Adam algorithm, the maximum number of iterations is 300 times, the initial learning rate is set to 0.005, the learning rate decay rate is set to 0.5, and the training will be carried out on the CPU.

[0089] The public power load dataset of a certain area in ISO-NE from 2012 to 2013 is used for simulation verification. This dataset includes data such as historical power load, historical meteorological information, and holidays. The sampling interval of the dataset is 1 h.

[0090] Cluster analysis is performed on the public power load dataset on a daily basis. The K-Medoids algorithm based on cosine similarity is used for cluster analysis, and the silhouette coefficient is used as the evaluation index. The maximum silhouette coefficient is 0.5112. When K = 3, the silhouette coefficient drops sharply, that is, the correlation within the cluster drops sharply. Therefore, K = 3 is the optimal number of clusters. To ensure the maximum difference between data in different clusters, when K = 3, cluster analysis is performed on the dataset. Through the analysis of the clustered dataset, it is found that: the first type of daily load curve has at most 364 days, mainly concentrated on non-high-temperature weekdays; the second type of daily load curve is less, with 182 days, mainly concentrated on high-temperature days; while the third type of daily load curve is also less, with 181 days, mainly concentrated on weekends / holiday dates. Therefore, according to the above clustering results, the daily load curves are divided into 3 typical scenarios: weekdays, high-temperature days, and weekends / holidays, as Figure 6 shown.

[0091] As can be seen from the above, the clustering center curves of weekdays, high-temperature days, and weekends / holidays are extracted from the dataset after K-Medoids clustering, and the JEP algorithm is used to extract and mark the extreme points of the load curves of 3 types of typical representative days. Because there are local extreme points and considering that the EIIR algorithm is to be used to calculate the importance of each point with respect to adjacent extreme values, but some extreme values are relatively close to each other, so 6 surrounding points are taken as the extreme value marking interval, which can include two extreme values and can better reflect the accuracy of importance. Based on the extreme points marked by the JEP algorithm, the EIIR algorithm is used to calculate the load time series curves of the typical representative days of the three scenarios of weekdays, high-temperature days, and weekends / holidays, and the importance of each point with respect to adjacent extreme points is extracted, such as Figure 7 shown.

[0092] To analyze the prediction performance of the present invention, the new feature sets combining the input features of external factors and the TAG sequence of each typical representative day are respectively input into the K-E-CNN-LSTM hybrid prediction model, K-CNN-LSTM model, K-CNN model, K-LSTM model, K-E-CNN model, and K-E-LSTM model to obtain the predicted loads of each typical representative prediction day.

[0093] For the weekday type load, the days to be measured are selected as October 21st to 25th, 2013. For the high-temperature type load, the days to be measured are selected as August 19th to 22nd, 2013. For the weekend / holiday type load, the days to be measured are selected as November 23rd to 24th, 2013. The load data of each type up to each prediction day are selected as the training set.

[0094] 1) Analysis of prediction results and average prediction errors

[0095] Under the 3 types of typical scenarios, the prediction results of the K-E-CNN-LSTM hybrid prediction model, K-CNN-LSTM model, K-CNN model, K-LSTM model, K-E-CNN model, and K-E-LSTM model are as Figure 8 shown. It can be seen that the K-E-CNN model and K-E-LSTM model with the EIIR algorithm are respectively superior to the K-CNN model and K-LSTM model without the EIIR algorithm, and the prediction results of the K-E-CNN model are basically the same as those of the K-CNN-LSTM prediction, indicating that adding the EIIR algorithm to the model can improve the prediction accuracy of the model.

[0096] From Figure 9The error radar chart shown can clearly display the MAPE (mean absolute percentage error), MAE (mean absolute error), and RMSE (root mean square error) of the prediction results under various models. It can be clearly seen that the prediction model with the smallest MAPE, MAE, and RMSE in the innermost circle is the K-E-CNN-LSTM hybrid prediction model proposed by the present invention. Followed by the K-E-CNN model using the EIIR algorithm and the K-CNN-LSTM model without using the EIIR algorithm.

[0097] Further compare the absolute errors of the K-E-CNN-LSTM hybrid prediction model, K-E-CNN model, and K-CNN-LSTM model. The absolute error bar chart is as Figure 10 shown, which can illustrate that the CNN-LSTM hybrid model has better prediction accuracy compared to the CNN and LSTM models.

[0098] In summary, the prediction result of the K-E-CNN-LSTM hybrid prediction model proposed by the present invention is closest to the true value. As shown in Table 3, the specific analysis is as follows.

[0099] For weekdays (October 21st to 25th), compared with the K-CNN-LSTM model and K-E-CNN model, the MAPE of the prediction result of the K-E-CNN-LSTM hybrid prediction model decreased by 31.1% and 4.84% respectively; the MAE decreased by 66.6425MW and 59.3209MW respectively; the RMSE decreased by 80.3985MW and 70.9573MW respectively.

[0100] For high-temperature days (August 19th to 22nd), compared with the K-CNN-LSTM model and K-E-CNN model, the MAPE of the prediction result of the K-E-CNN-LSTM hybrid prediction model decreased by 19.14% and 9.14% respectively; the MAE decreased by 69.0304MW and 23.1166MW respectively; the RMSE decreased by 80.0319MW and 35.3514MW respectively.

[0101] For weekends / holidays (November 23rd to 24th), compared with the K-CNN-LSTM and K-E-CNN models, the MAPE of the prediction result of the K-E-CNN-LSTM hybrid prediction model decreased by 23.88% and 30.14% respectively; the MAE decreased by 69.6687MW and 97.1494MW respectively; the RMSE decreased by 64.811MW and 81.0642MW respectively.

[0102] Table 3 Load Forecasting Error

[0103]

[0104] 2) Analysis of Prediction Error in Extreme Point Interval

[0105] As can be seen from the above, the prediction effects of the K-E-CNN-LSTM hybrid prediction model, K-CNN-LSTM model, and K-E-CNN model among the six models are significantly better than the other three. Therefore, the K-E-CNN-LSTM hybrid prediction model, K-CNN-LSTM model, and K-E-CNN model are used to discuss the prediction error in the extreme point interval. Take three points, namely the extreme points of each type of typical representative day as a single point, as the extreme point interval. As shown in Table 4.

[0106] Table 4 Extreme Point Interval

[0107] Load type Extreme point interval Weekday 3 - 5, 11 - 13, 15 - 17, 19 - 21 High temperature day 3 - 5, 16 - 18, 19 - 21, 20 - 22 Weekend / holiday 3 - 5, 11 - 13, 15 - 17, 18 - 20

[0108] Similarly, calculate the prediction accuracy of the extreme value intervals of each predicted representative day for the K-E-CNN-LSTM hybrid prediction model, K-CNN-LSTM model, and K-E-CNN model. The results are shown in Tables 5 and 6. For the extreme point intervals on weekdays (from October 21st to 25th), the prediction results of the K-E-CNN-LSTM hybrid prediction model and K-E-CNN model are both better than the K-CNN-LSTM model. Compared with the K-CNN-LSTM model, the MAPE, MAE, and RMSE of the K-E-CNN-LSTM hybrid prediction model at the 4th moment decreased by 47.06%, 118.149MW, and 108.575MW respectively. At the 12th moment, the MAPE, MAE, and RMSE decreased by 29.51%, 51.5455MW, and 41.6209MW respectively. At the 16th moment, the MAPE, MAE, and RMSE decreased by 51.2%, 93.267MW, and 91.4629MW respectively. At the 20th moment, the MAPE, MAE, and RMSE decreased by 12.73%, 33.0749MW, and 45.0718MW respectively.

[0109] For the extreme value point interval of high temperature days (from August 19th to 22nd), the prediction results of the K-E-CNN-LSTM prediction model are better than those of the K-CNN-LSTM model and the K-E-CNN model. Compared with the K-CNN-LSTM model, the MAPE, MAE, and RMSE of the K-E-CNN-LSTM prediction model at the 4th moment decreased by 75.69%, 232.049MW, and 219.9413MW respectively. At the 17th moment, the MAPE, MAE, and RMSE decreased by 1.43%, 15.0819MW, and 14.7907MW respectively. At the 20th moment, the MAPE, MAE, and RMSE decreased by 22.97%, 90.2671MW, and 72.1902MW respectively. At the 21st moment, the MAPE, MAE, and RMSE decreased by 38.08%, 173.0539MW, and 189.2567MW respectively.

[0110] For the extreme value point interval of weekends / holidays (from November 23rd to 24th), the prediction results of the K-E-CNN-LSTM prediction model are better than those of the K-CNN-LSTM model and the K-E-CNN model. Compared with the K-CNN-LSTM model, the MAPE, MAE, and RMSE of the K-E-CNN-LSTM prediction model at the 4th moment decreased by 19.08%, 111.5622MW, and 112.285MW respectively. At the 12th moment, the MAPE, MAE, and RMSE decreased by 53.33%, 21.3503MW, and 27.275MW respectively. At the 16th moment, the MAPE, MAE, and RMSE decreased by 27.11%, 67.5884MW, and 94.0494MW respectively. At the 19th moment, the MAPE, MAE, and RMSE decreased by 30.26%, 38.2105MW, and 118.0402MW respectively.

[0111] Table 5 Prediction of extreme value point interval for working days (from October 21st to 25th)

[0112]

[0113] Table 6 Prediction of extreme value point interval for high temperature days (from August 19th to 22nd)

[0114]

[0115] Table 6 Prediction of extreme value point interval for weekends / holidays (from November 23rd to 24th)

[0116]

[0117] The present invention effectively improves the load prediction accuracy in the extreme value point interval, thereby improving the overall load prediction accuracy.

[0118] It can be proved hereby that:

[0119] The K-Medoids clustering algorithm based on cosine similarity adopted by the present invention can clearly, explicitly and reasonably divide typical scenarios according to the daily load curve.

[0120] Extracting the extreme point importance feature sequence of the daily load curve based on the EIIR algorithm can represent the degree of importance change with digital marks, especially for the extreme point moments. To a certain extent, it reduces the influence of data redundancy on the prediction model, enriches the input of the prediction model, can greatly improve the prediction accuracy of the prediction model, and has a certain universality.

[0121] The K-E-CNN-LSTM hybrid prediction model proposed in this paper has significantly better prediction accuracy in the extreme point interval than the K-CNN-LSTM model without extracting importance.

[0122] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A short-term load forecasting method based on time series importance analysis and feature extraction, characterized in that: The following steps are involved: S1. Perform cluster analysis on the power load data set containing historical load data, historical meteorological data and date data on a daily basis, and based on the clustering results, use each cluster center as a typical representative daily load curve; S2. extracting the morphological characteristics of the time series data corresponding to each type of typical representative daily load curve through an extreme point importance extraction algorithm, and obtaining an extreme point importance characteristic sequence of each type of typical representative daily load curve; S3. Based on the historical meteorological data and date data in the power data set, external factor input features are selected, and the external factor input features are reconstructed into a new feature set with the extreme point importance feature sequence of each type of typical representative daily load curve. The new feature set of each type of load is input into the CNN-LSTM hybrid model for prediction to obtain the predicted load of each type of prediction day.

2. The short-term load forecasting method based on time series importance analysis and feature extraction according to claim 1 is characterized in that: The extreme point importance extraction algorithm consists of two parts: an extreme point judgment algorithm, which is used to judge whether an element in a sequence is an extreme point and mark it; an importance marking algorithm, which is used to mark the importance of each element in the sequence according to the level and output an extreme point importance feature sequence.

3. The short-term load forecasting method based on time series importance analysis and feature extraction according to claim 1 or 2, characterized in that: The method for obtaining the importance characteristic sequence of the extreme points of each type of typical representative daily load curve is as follows: in the process of traversing the typical representative daily load curve using the importance marking algorithm according to the marking, in each cycle, the extreme points of the typical representative daily load curve are first screened and marked using the extreme point judgment algorithm, and then the importance level is judged according to the number of markings to extract the importance characteristic sequence of the extreme points of the typical representative daily load curve.

4. The short-term load forecasting method based on time series importance analysis and feature extraction according to claim 3 is characterized in that: The specific method of first using the extreme point judgment algorithm to screen and mark the extreme points of the typical representative daily load curve is as follows: Suppose a sequence S = {s1,s2,s3,s4,…,s n }, the sequence length is n, and the mth element of sequence S is s m , v represents the current computational neighborhood; if the element s m Than element s m The first v elements and the element s m If the following v elements are all large, then element s m is marked as 1; if the element s m Than element s m The first v elements and the element s m If the following v elements are all small, then element s m is marked with -1, otherwise the element s m The sign of is taken as 0, and the formula is as follows: Among them, Flag is the extreme value mark, and k is the calculation neighborhood range of the extreme value.

5. The short-term load forecasting method based on time series importance analysis and feature extraction according to claim 3 is characterized in that: The method for extracting the importance characteristic sequence of the extreme value points of the typical representative daily load curve is: For the sequence S = {s1, s2, s3, s4, ..., s n }, the importance mark algorithm is calculated within the range of the k elements before and after the neighborhood: first, the extreme point importance feature sequence is initialized, and then the sequence S is traversed through the inner and outer loops; the outer loop is from 1 to k, and the inner loop traverses each element of the sequence S; in each inner loop, the degree of association between each point and the extreme point is judged according to the mark of the extreme point judgment algorithm and the value of the extreme point importance feature sequence. If the preset conditions are met, the value of the extreme point importance feature sequence is updated accordingly; After completing the inner and outer loops, the updated extreme point importance feature sequence is obtained; extreme point importance feature sequence TAG = {z1, z2, z3, …, z n }, where z m Indicates the importance of the mth element in the S sequence.

6. The short-term load forecasting method based on time series importance analysis and feature extraction according to claim 1 or 5, characterized in that: The K-Medoids algorithm is used for cluster analysis. The cosine similarity is selected as the measurement index, and the silhouette coefficient is used as the evaluation index. The K value with the highest overall silhouette coefficient is selected as the optimal cluster number.

7. The short-term load forecasting method based on time series importance analysis and feature extraction according to claim 6 is characterized in that: The cosine similarity is calculated as follows: Let x i =[x i1 , x i2 , …, x in ] and x j =[x j1 , x j2 , …, x jn ] are the load data sequences of the i-th and j-th day of a certain area, respectively. The calculation formula of the cosine similarity of the load data sequences of the i-th and j-th day is as follows: Among them, S cos (x i ,x j ) is the cosine similarity of the load data series between the i-th day and the j-th day, x iq For the sequence x i The qth element in x iq For the sequence x j The qth element in , where n is the length of the sequence.

8. The short-term load forecasting method based on time series importance analysis and feature extraction according to claim 7 is characterized in that: The overall silhouette coefficient is calculated as follows: First calculate the sequence x i and the sequence x j The silhouette coefficient is calculated as follows: In the formula, a(i) is the sequence x i To other sequences of the same type x j The average cosine similarity of; b(i) is the sequence x i To other classes sequence x j The average cosine similarity of Next, the silhouette coefficients of all sequences are averaged to obtain the overall silhouette coefficient of the clustering results. The formula is as follows: In the formula, the overall silhouette coefficient S C The value of is [-1,1], and N is the total number of sequences.

9. The short-term load forecasting method based on time series importance analysis and feature extraction according to claim 1 or 8, characterized in that: The specific structure of the CNN-LSTM hybrid model includes the following layers connected in sequence: convolution layer I, maximum pooling layer I, convolution layer II, maximum pooling layer II, convolution layer III, maximum pooling layer III, flattening layer, LSTM layer I, random dropout layer I, LSTM layer II, random dropout layer II and fully connected layer.

10. The short-term load forecasting method based on time series importance analysis and feature extraction according to claim 9 is characterized in that: The external factor input features include temperature, humidity, temperature and humidity index and date type.

Images