WO2018045642A1 - 一种母线负荷预测方法 - Google Patents

一种母线负荷预测方法 Download PDF

Info

Publication number
WO2018045642A1
WO2018045642A1 PCT/CN2016/108051 CN2016108051W WO2018045642A1 WO 2018045642 A1 WO2018045642 A1 WO 2018045642A1 CN 2016108051 W CN2016108051 W CN 2016108051W WO 2018045642 A1 WO2018045642 A1 WO 2018045642A1
Authority
WO
WIPO (PCT)
Prior art keywords
load
data
bus
sample
value
Prior art date
Application number
PCT/CN2016/108051
Other languages
English (en)
French (fr)
Inventor
孟强
王一蓉
郝悍勇
张建
杜朝晖
吴润泽
邓伟
杨松楠
范军丽
包正睿
Original Assignee
国网山西省电力公司晋城供电公司
国家电网公司
北京国电通网络技术有限公司
华北电力大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国网山西省电力公司晋城供电公司, 国家电网公司, 北京国电通网络技术有限公司, 华北电力大学 filed Critical 国网山西省电力公司晋城供电公司
Priority to AU2016325186A priority Critical patent/AU2016325186A1/en
Publication of WO2018045642A1 publication Critical patent/WO2018045642A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • G06N20/20Ensemble learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/23Clustering techniques
    • G06F18/232Non-hierarchical techniques
    • G06F18/2321Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133Distances to prototypes

Definitions

  • the invention relates to the technical field of power system engineering, in particular to a method for predicting bus load.
  • Bus load forecasting is an important part of power system planning and the basis of power system economic operation. Its prediction results can better achieve decentralized load management, directly affect the analysis results of power grid follow-up safety check, and estimate the dynamic state of power grid. , reactive power optimization, local control of plant sites and reducing power generation costs are of great significance.
  • the bus load As the substation load of the substation, the bus load is small, the stability is not strong, the change trend is not obvious, and the high frequency fluctuation component has become a difficult point to improve the prediction accuracy.
  • Short-term bus load forecasting generally refers to real-time forecasting. It requires not only high prediction accuracy but also fast calculation speed. Due to the characteristics of bus load itself and the inherent defects of traditional forecasting methods, there are few studies relative to system load forecasting.
  • Bus load forecasting generally uses regression analysis, neural network and support vector regression. Different prediction methods extract historical load information from different angles, and the requirements for sample data size and complexity are different. When dealing with the bus load prediction problem considering multiple influencing factors, the regression analysis method is simple to implement, but it has high requirements on the stability of the sample, and it is not suitable to directly deal with the bus load with strong volatility. The neural network and support vector regression method have strong nonlinear fitting ability, and the requirements for sample stability are not high. However, when the input data has too many dimensions and the sample size is large, the training of the model is easy to fall into local optimum. And high complexity, which reduces prediction accuracy and computational efficiency.
  • the object of the present invention is to propose a bus load prediction method capable of analyzing the variation law of the bus load, improving the prediction accuracy of the bus load, and satisfying the real-time requirement.
  • the bus load prediction method provided by the present invention based on the above object includes:
  • the key influence factor value of the busbar to be tested is used as a feature vector, and is input into a pre-built classification model to obtain a category of the date to be measured, and the pre-built classification model is aggregated according to the key influencing factors and the historical load data of the busbar to be tested.
  • the corresponding multiple regression model is selected from the pre-trained multiple regression prediction model to predict the load of the bus to be tested; the multi-regressive prediction model obtained by the pre-training is based on the historical load of the bus to be tested.
  • the clustering analysis results of the data for different typical load patterns, select the corresponding historical load data and key influencing factor values as the prediction model trained by the sample data.
  • the method for constructing the classification model is:
  • the horizontal comparison method is used to screen out the abnormal data in the historical load data of the bus to be tested, and the abnormal data is corrected according to a preset algorithm to obtain corrected load data;
  • the gray correlation projection method is used to calculate the degree of correlation between the influencing factors and the historical load data. From the large to the small, the preset number of the influencing factors corresponding to the degree of association is selected as the key influencing factors. ;
  • the historical load data of the measured busbars is clustered and analyzed, and different typical load patterns are obtained.
  • a random forest algorithm is used to construct the classification model, and the mapping relationship between key influencing factors and various bus loades is established.
  • the using the horizontal comparison method to filter the abnormal data in the historical load data of the bus to be tested includes:
  • the historical load data of the measured bus bar is cleaned by the horizontal comparison method to identify abnormal data; wherein the horizontal comparison method uses the rate of change of the adjacent point load as a criterion for judging whether it is abnormal data, and the load change rate calculation formula as follows:
  • ⁇ (d,t) is the rate of change of the load of the busbar to be tested at time t on the dth day
  • L(d,t) is the load value of the busbar to be tested at time t on the dth day; whether the data of the sampled point is abnormal data
  • the load value at the current time is compared with the load value at the previous time in chronological order, and is corrected immediately when the abnormal data is detected, and the load data is corrected as the comparison value of the load data at the next time.
  • the method for modifying the abnormal data according to a preset algorithm is:
  • the filtered abnormal data is corrected according to the following algorithm.
  • the specific algorithm formula is:
  • ⁇ m is a weight coefficient, which is used to indicate the degree of influence of the load on the d-day t time load on the d-day t-time load; ⁇ is the smoothing coefficient, and L(dm, t) is the test bus at the d-th day t time. Load value.
  • the gray correlation projection method is used to calculate the degree of association between the influencing factor and the historical load data, and the influencing factors corresponding to the preset degree are selected as the key influencing factors, including:
  • n denotes the number of comparison sequences
  • X 0 represents a load sequence or a reference sequence corresponding to the modified load data
  • X i represents an i-th comparison sequence
  • the correlation coefficient ⁇ 0i (k) of the normalized comparison sequence e i and the kth element of the normalized reference sequence e 0 is calculated as:
  • is the two-level minimum difference
  • is the two-level maximum difference
  • is the resolution coefficient
  • the gray correlation coefficient judgment matrix F is calculated as follows:
  • ⁇ 00 (j) is the correlation coefficient between the normalized reference sequence e 0 and its own jth element, and the value is 1;
  • the entropy weight method is used to assign the weight of the column vector of the matrix E. Except for the first row element, the calculation formula is as follows:
  • E j represents the information entropy of the data of the jth column
  • p ij represents the proportion of the value of the element e ij in the column element
  • w j represents the weight of the data of the jth column
  • W (w 1 ,. ..,w j ,...,w m ).
  • w j ⁇ 0i (j) is the weighted correlation coefficient of the jth element of the sequence e i and e 0 ;
  • D i representative of the comparison of sequences X projection value is associated with a test bus load of i, D i between 0 and 1;
  • the M factors with the highest projection values are selected as the key influencing factors.
  • the calculating step of performing cluster analysis on the historical load data of the bus to be tested according to the preset K-means clustering algorithm is:
  • n indicates that each load sequence includes n+1 data points, k represents the number of centroids, and x it represents the load value at time t+1 of the i-th load sequence. Representing the t+1th element value in the first centroid vector;
  • centroid of each class is updated; the number of samples of the jth class is m j , and the sample set is S j , and the centroid vector of the sample is recalculated as:
  • the determining whether the preset termination condition is reached is: after each iteration is completed, calculating a distance interval before and after each type of centroid update according to the following formula:
  • the step of constructing the classification model by using a random forest algorithm includes:
  • N influencing factors are randomly selected as feature attributes of each training sample set; wherein, N ⁇ M;
  • the s sample sets TS i are respectively trained to generate the corresponding decision tree Tree i ; during the whole forest growth process, the feature attributes of the decision tree Tree i remain unchanged; except for the leaf nodes, each decision tree adopts the CART algorithm.
  • the Gini index serves as the basis for node splitting:
  • d is the current node
  • p i,d is the proportion of the target class i
  • k is the number of target classes
  • the attribute with the smallest Gini index is selected as the split attribute a;
  • the Gini index of the current node d according to the split attribute a is:
  • Gini(a,d) p L Gini(d L )+p R Gini(d R ), where d L and d R are left and right child nodes of node d, and p L and p R are left and right child nodes The proportion of the sample set in the parent node;
  • the decision tree grows, and there are s trees.
  • Each tree does not need pruning operation, and its leaf nodes correspond to cluster analysis.
  • a certain clustering result; the s decision tree is combined to obtain a random forest classification model; when a test sample is input, each decision tree is used to classify it, and s classification results are obtained, and the highest proportion is selected.
  • the class serves as the category for this test sample.
  • the multiple regression prediction model is established by multiple linear regression, and the regression parameters are determined by a least squares method.
  • b 0 is the regression constant term
  • b 1 , b 2 ,..., b n is the regression coefficient
  • f(x i ) is the regression prediction value
  • x i represents an input sequence of samples X i, x in the n-th value of influencing factors, y i is the true value of the bus load corresponding to the output;
  • the input matrix X, the output matrix Y, and the coefficient matrix B are as follows:
  • the least squares method is used to determine the parameter estimates of the regression equation:
  • the bus load prediction method determines the key influencing factors of the load data by the gray correlation projection method, and adopts the K-means clustering and the random forest classification method, and combines the regression analysis method based on various loads.
  • the characteristic laws of the sequence are respectively established statistical prediction model, the internal law of the bus load data is mined, the complex historical load is decomposed into several typical load patterns, and the targeted training prediction model is selected, so that the matching is selected according to the category of the day to be tested.
  • the model completes the prediction.
  • the bus load prediction method can not only improve the accuracy of the bus load prediction, but also meet the real-time requirements of the prediction, so that the prediction result is more stable and reliable.
  • FIG. 1 is a flow chart of an embodiment of a bus load prediction method provided by the present invention.
  • FIG. 2 is a flow chart of another embodiment of a bus load prediction method provided by the present invention.
  • FIG. 3 is a flow chart of an embodiment of a gray correlation projection method provided by the present invention.
  • FIG. 4 is a flowchart of an embodiment of a preset K-means clustering algorithm provided by the present invention.
  • FIG. 5 is a clustering result diagram of load data in a bus load prediction method provided by the present invention.
  • FIG. 6 is a flowchart of an embodiment of a random forest algorithm provided by the present invention.
  • FIG. 7 is a thumbnail diagram of a decision tree generated by the random forest algorithm provided by the present invention.
  • Fig. 8 is a graph showing the result of bus load prediction in the bus load prediction method provided by the present invention.
  • the present invention introduces data mining methods in big data technology to obtain several types of bus load patterns, and combines regression analysis methods to establish statistics based on the characteristic laws of various load sequences.
  • the model is predicted so that the corresponding model is selected according to the category of the day to be measured to complete the prediction, that is, the model matching is implemented.
  • the method can better mine the variation law of the bus load by time domain decomposition of the historical load sequence, and effectively improve the prediction accuracy.
  • the prediction model library can be directly obtained. The load of the bus at each moment of the day is predicted, and it is not necessary to retrain the model before each prediction to meet the real-time requirements of the short-term prediction.
  • a bus load prediction method including:
  • Step 101 The key influence factor value of the current busbar to be tested is used as a feature vector, and is input into a pre-built classification model to obtain a category of the date to be measured; wherein the pre-built classification model is based on the key influencing factor and the busbar to be measured.
  • the historical load data is the result of cluster analysis, and the mapping relationship between the key influencing factors and the bus load is established;
  • Step 102 Select a corresponding multiple regression prediction model from the pre-trained multiple regression prediction model according to the type of the day to be measured, and perform prediction of the bus load to be tested; wherein the pre-trained multiple regression prediction model is based on The clustering analysis results of the historical load data of the measured busbars are selected for the corresponding typical load patterns, and the corresponding historical load data and key influencing factor values are selected as the prediction models trained by the sample data.
  • the bus load prediction method of the present invention achieves efficient and accurate prediction of the bus load based on model matching.
  • a method for constructing a classification model including:
  • the horizontal comparison method is used to screen out the abnormal data in the historical load data of the bus to be tested, and the abnormal data is corrected according to a preset algorithm to obtain corrected load data;
  • the gray correlation projection method is used to calculate the degree of correlation between the influencing factors and the historical load data. From the large to the small, the preset number of the influencing factors corresponding to the degree of association is selected as the key influencing factors. ;
  • the historical load data of the measured busbars is clustered and analyzed, and different typical load patterns are obtained.
  • a random forest algorithm is used to construct the classification model, and the mapping relationship between key influencing factors and various bus loades is established.
  • the algorithm selected in the above construction method is only an optional manner in the embodiment of the present invention, and other suitable algorithms may be selected to construct the model.
  • the order between the above steps is not limited, and only the logical relationship needs to be made to conform to the normal algorithm steps.
  • the bus load prediction method includes:
  • Step 201 Acquire and store historical load data of the bus to be tested and impact factor data related to the load; wherein the influencing factors include: weather information, day type, and the like, which can affect the bus load.
  • Step 202 The horizontal comparison method is used to screen out the abnormal data in the historical load data of the bus to be tested, and the abnormal data is corrected according to a preset algorithm to obtain corrected load data; that is, the bus load historical data is preprocessed, and the horizontal load is utilized.
  • the comparison method processes the historical load data of the measured busbars, identifies outliers in the load sequence, and corrects them.
  • Step 203 combined with the modified load data and the load-related influencing factors, the gray correlation projection method is used to calculate the degree of correlation between the influencing factors and the historical load data, and the influencing factors corresponding to the preset degree are selected from large to small. Key influencing factors; the predictive model has a reasonable input dimension by determining the key factors affecting the bus load.
  • Step 204 Perform cluster analysis on the historical load data of the bus to be measured according to the preset K-means clustering algorithm, so that the initial cluster centroid is evenly distributed in the sample space, and the historical load data with similar characteristics are classified into one class. Different typical load patterns are obtained. Among them, the bus load clustering analysis specifically analyzes the historical load data of the measured busbars in units of days, and classifies the bus loads with similar characteristics into one class, and obtains some typical load patterns. Based on the preset K-means clustering algorithm, the initial cluster centroid is evenly distributed in the sample space, which improves the convergence speed and clustering accuracy of the algorithm.
  • Step 205 According to the result of the key influencing factors and cluster analysis, construct a classification model by using a random forest algorithm, and establish a mapping relationship between key influencing factors and various bus loades;
  • Step 206 According to the cluster analysis result of the historical load data of the bus to be tested, select corresponding historical load data and key influencing factor values as sample data for different typical load patterns, and obtain different multi-regressive prediction models.
  • Step 207 Taking the value of the key influencing factor of the bus to be tested as the feature vector, inputting into the classification model, obtaining the category of the day to be measured, and selecting a corresponding multiple regression prediction model according to the type of the day to be tested, and performing the bus load to be tested. Prediction.
  • the bus load prediction method provided by the present invention differs in the historical load sequence of the bus to be measured.
  • the constant value is corrected, and the key influencing factors of the bus load are determined by the grey correlation projection method.
  • the improved K-means algorithm is used to cluster the revised historical data, and the random forest algorithm is used to construct the classification model to establish the clustering result and key.
  • the bus load prediction method not only improves the accuracy of the bus load prediction, but also satisfies the real-time requirements of the prediction, so that the prediction result is more stable and reliable.
  • the screening, by using the horizontal comparison method, the abnormal data in the historical load data of the bus to be tested includes: cleaning the load history data of the bus to be measured by the horizontal comparison method, and identifying and correcting the abnormality. value.
  • the horizontal comparison method uses the rate of change of the adjacent point load as the criterion for judging abnormal data.
  • the calculation formula of the load change rate is as follows:
  • ⁇ (d,t) is the rate of change of the bus load at time t on day d
  • L(d,t) is the load value of the bus to be tested at time t on day d
  • the method for judging whether the sampling point data is an abnormal value is as follows:
  • i represents the number of days from the current day d
  • L(di, t-1) is the sample point data at the time t-1 of the di day, and there are q every day. +1 sample point.
  • the identified abnormal data is corrected according to the following formula:
  • ⁇ m is a weight coefficient, which is used to indicate the degree of influence of the load on the d-day t time load on the d-day t-time load
  • is the smoothing coefficient
  • L(dm, t) represents the bus load value at the d-th day t time. That is, the data of the last m days is used as the correction data.
  • the whole process is performed in chronological order, only compared with the load value of the previous moment.
  • the abnormal data is detected, it is corrected immediately, and as the comparison value of the next time data, all the historical load data is corrected.
  • the gray correlation projection method is used to calculate the image.
  • the degree of correlation between the response factor and the historical load data selects the preset number of the influencing factors corresponding to the degree of association as the key influencing factors, including: using gray correlation projection method to calculate the correlation degree between each influencing factor and the bus load. .
  • the calculated gray correlation projection values are sorted, and the M factors with larger values are selected as the key influencing factors.
  • the specific calculation steps are as follows:
  • the reference sequence and the comparison sequence are first determined.
  • n represents the number of comparison sequences.
  • X 0 represents a reference sequence, that is, a load sequence
  • X i represents an i-th comparison sequence, that is, an influence factor sequence.
  • step 302 the sequence matrix X is normalized according to the following formula:
  • e ij represents a normalized value of the element x ij in the matrix X, and the value of e ij is generally limited to between 0.1 and 1.
  • e 0 and e i are normalized reference sequences and comparison sequences.
  • step 303 the correlation coefficient is calculated.
  • the correlation coefficient ⁇ 0i (k) of the normalized comparison sequence e i and the kth element of the reference sequence e i is determined by:
  • is a two-level minimum difference
  • is a resolution coefficient, and preferably, ⁇ takes a value of 0.5.
  • the gray correlation coefficient judgment matrix F is obtained from the correlation coefficient as follows:
  • ⁇ 00 (j) represents the correlation coefficient between the normalized load sequence e 0 and its own j-th element, so the value is 1.
  • step 304 the entropy weight method is used to assign a weight to the column vector of the matrix E (except the first row element), and the calculation formula is as follows:
  • E j represents the information entropy of the data of the jth column
  • p ij represents the proportion of the value of the element e ij in the column element
  • w j represents the weight of the data of the jth column
  • W (w 1 ,. ..,w j ,...,w m ).
  • Step 305 by step 304, a weighted gray correlation judgment matrix F' is obtained.
  • w j ⁇ 0i (j) is the weighted correlation coefficient of the jth element of the sequence e i and e 0 .
  • Step 306 calculating a gray relational projection value, and the calculation formula is as follows:
  • X D i representative of the comparison of sequences associated projection values of i and the bus load, ranging between 0 and 1, the value closer to 1, the greater the degree of correlation; values closer to 0, the weaker the degree of correlation.
  • the gray correlation projection value can more comprehensively reflect the similarity of the development trend of the two objects.
  • Step 307 finally selecting the M influencing factors with the largest projection value as the key influencing factors.
  • the M value can be set as needed, and the selected process can first sort all the projection values, and select the M factors of the top projection value from large to small. For key influencing factors, sorting is also not required.
  • the calculating step of performing cluster analysis on the historical load data of the bus to be tested according to the preset K-means clustering algorithm includes:
  • the modified historical load data is used as a sample set, and the improved K-means algorithm is used for bus load characteristic clustering.
  • the specific calculation process is as follows:
  • Step 401 initial clustering the selection of the centroid.
  • the initial clustering centroid of the traditional algorithm is randomly selected, and the improved algorithm proposed by the present invention uniformly distributes the initial centroid in the sample space by using the distance as a standard, thereby effectively reducing the influence of the initial centroid on the clustering result.
  • the sample set X ⁇ x 1 ,...,x i ,...,x m ⁇
  • the load sequence x i ⁇ x i0 ,...,x it ,...,x in ⁇
  • initial centroid set A sample is randomly selected as the first centroid.
  • Step 402 Calculate the Euclidean distance d from the first centroid for each of the remaining samples by the following formula:
  • n denotes that each load sequence includes n+1 data points
  • k represents the number of centroids, that is, there are k load patterns.
  • x it represents the load value at time t+1 in the i-th load sequence, Represents the initial centroid vector of class 1. The t+1th element value in the middle.
  • Step 403 Determine whether k centroids have been selected. If k centroids have been selected, it indicates that the screening operation has been completed, and the next step can be performed. If the number of centroids has not reached k, continue to select the centroid.
  • step 404 the obtained k centroids respectively correspond to different types of load patterns, or a corresponding relationship is established, and the number of iterations is recorded as 0.
  • Step 405 sample assignment or sample determination.
  • the Euclidean distance of each sample and all centroids is calculated separately, and the load type corresponding to the centroid of the sample having the smallest Euclidean distance is selected, and the sample is assigned to the load type. Calculated as follows:
  • the centroid is updated. Calculate the centroid vector for each class and start updating the centroid of each class after all samples have been assigned. Let the number of samples of the jth class be m j and the sample set thereof be S j , and recalculate the centroid vector of the class as:
  • step 407 it is determined whether the termination condition is reached.
  • the algorithm determines whether the algorithm can be terminated by defining the difference tolerance ⁇ . After each iteration is completed, calculate the distance interval before and after each type of centroid update according to the following formula:
  • the historical load of the busbar is classified into the k-type load mode. Under each load mode, the change trend of the bus load is more obvious and the regularity is stronger, which lays a foundation for constructing the model and improving the prediction accuracy.
  • step 408 k kinds of clustering results are obtained.
  • the step of constructing a classification model by using a random forest algorithm further includes:
  • N (N ⁇ M) influencing factors are randomly selected as the characteristic attributes of each training sample set
  • each decision tree uses the Gini index in the CART algorithm as the basis for node splitting:
  • d is the current node
  • p i,d is the proportion of the target class i
  • k is the number of target classes. The closer the Gini index is to 0, the higher the purity and the better the effect.
  • the Gini index of the current node d divided by the split attribute a is:
  • Gini(a,d) p L Gini(d L )+p R Gini(d R ), where d L and d R are left and right child nodes of node d, and p L and p R are left and right child nodes The proportion of the sample set in the parent node.
  • a random forest classification model is obtained, which establishes the mapping relationship between the influencing factors and the clustering results, and determines the classification of the days to be tested in real time and effectively, which lays a foundation for the prediction model matching.
  • the step of the training obtaining different multiple regression prediction models further includes:
  • the present invention selects multiple linear regression model for prediction.
  • Each type of bus load data and key influencing factors are used as training samples to construct several prediction models.
  • the prediction model was established using multiple linear regression and the parameters were determined by least squares.
  • x i represents the input sequence of the sample X i
  • x in is the nth influencing factor value
  • y i is the actual value of the bus load corresponding to the output.
  • b 0 is a regression constant term
  • b 1 , b 2 , ..., b n is a regression coefficient
  • f(x i ) is a regression prediction value
  • the input matrix X, the output matrix Y, and the coefficient matrix B are as follows:
  • the least squares method is used to determine the parameter estimates of the regression equation:
  • the key influencing factors at each moment of the day to be measured are input into the random forest classification model, and the categories of the days to be measured are obtained, thereby selecting the corresponding multiple regression prediction model to complete the load prediction.
  • the bus load prediction method of the invention has the following effects: through the cluster analysis of the historical bus load, the establishment of the statistical model and the model matching of the day to be tested, the variation law of the bus load is effectively excavated, and a more accurate prediction result is obtained;
  • the pre-predictive training model further satisfies the real-time requirements of bus load forecasting.
  • a 110kV bus active load of a certain grid city company is used as a prediction example.
  • the first stage is the preprocessing part of the historical load data.
  • the load value of the 110kV busbar of a substation is selected as the forecasting object at 24 o'clock, and the load data of the forecast year is used as the training sample.
  • the data format is as shown in Table 1:
  • Each row of data in Table 1 represents a 24-dimensional load sample sequence. Due to the instability of the acquisition system, there may be some abnormal values in the historical load data, that is, abnormal data, which will seriously affect the prediction accuracy.
  • the anomaly data is corrected by the horizontal comparison method.
  • the sampling period is set to 1 hour, a total of 24 data points per day, and the rate of change of the load at the adjacent time is sequentially calculated in chronological order:
  • the calculated maximum value of the load change rate within 7 days before the sampling point at each time is used as a criterion for whether the data at the sampling time is an abnormal value, and the judgment formula is as follows:
  • the correction formula is as follows:
  • the second phase is the stage of identification of key influencing factors.
  • the influencing factors are: maximum temperature (T max ), minimum temperature (T min ), real-time temperature (RT), average wind speed (AW), relative humidity (RH). Average precipitation (AP), day type (DT) and season type (ST).
  • T max maximum temperature
  • T min minimum temperature
  • RT real-time temperature
  • AW average wind speed
  • RH relative humidity
  • AP average precipitation
  • DT day type
  • ST season type
  • the column data corresponding to each influencing factor is the comparison sequence X i
  • the load data at each sampling time is the reference sequence X 0 .
  • the data is the actual measured value.
  • comparison sequence and the reference sequence are respectively standardized by the formula, and the corresponding values are limited to between 0.1 and 1, and the standardized sequence matrix E is obtained:
  • e nm represents a normalized value of the element x nm in the sequence matrix X.
  • the gray correlation coefficient between the elements of the comparison sequence and the reference sequence is obtained by the correlation coefficient calculation formula, thereby obtaining the gray correlation coefficient judgment matrix F.
  • a weighted gray correlation judgment matrix F' can be obtained :
  • ⁇ 0i (j) is the correlation coefficient between the normalized comparison sequence e i and the jth element of the reference sequence e 0
  • w j is the weight of the column vector j
  • w j ⁇ 0i (j) is the sequence e i and e 0
  • the gray correlation projection value D i of the reference sequence X i and the comparison sequence X 0 is calculated as follows:
  • Sorting selects the highest temperature, real-time temperature, relative humidity, minimum temperature and day type as the key influencing factors affecting the load trend of the busbar to be tested.
  • the third stage is the historical load clustering stage of the bus to be tested.
  • a specific implementation process for the improved K-means algorithm is presented.
  • the historical load sample set X has a total of 365 sample sequences, and a sample sequence is randomly selected as the first initial centroid vector. Calculate each of the remaining samples separately Euclidean distance d:
  • centroid vector After each assignment of the sample, the centroid of each type of load sample needs to be updated. If the class j sample number is m j and its sample set is S j , then the centroid vector Elements in Updated to:
  • Class 1 contains 82 days, mainly concentrated on rest days.
  • Class 2 and Class 5 mainly contain 112 and 89 working days. The difference between these two types of load values is mainly due to seasonal factors.
  • Class 3 contains 25 days and is subject to extremes. Due to the weather, the load is much higher than other categories; class 4 mainly contains all the holidays, the shape of the curve is obviously different from other classes; class 6 contains 31 days, and the load curve fluctuates greatly.
  • the fourth stage is the construction phase of the random forest classification model. Referring to Figure 6, the specific implementation process of the random forest algorithm.
  • the data of the training sample set TS i is shown in Table 4:
  • Three influencing factors are randomly selected as feature attributes for each training sample set.
  • each decision tree uses the Gini index in the CART algorithm as the basis for node splitting:
  • the Gini index of the current node d by attribute a is:
  • Gini(a,d) p L Gini(d L )+p R Gini(d R ), where d L and d R are left and right child nodes of node d, and p L and p R are left and right child nodes The proportion of the sample set in the parent node.
  • Figure 7 is a thumbnail of a decision tree generated in a random forest classification model.
  • the fifth stage is the training phase of the predictive model. According to the clustering result, the bus load data and key influencing factors values in each load mode are integrated into the training sample set, as shown in Table 5:
  • m is the number of samples for each type of sample set.
  • x i represents the sequence of influencing factors of the sample X i
  • y i is the actual value of the bus load corresponding to the output.
  • Each type of load pattern needs to construct a prediction model based on its own training sample set, and finally obtain six regression equations.
  • the multiple regression equation is obtained by:
  • Equation 1 4.9 0.64 0.51 0.36 0.32 0.23 Equation 2 -8.5 0.57 0.60 0.41 0.26 0.18 Equation 3 6.5 0.48 0.66 0.35 0.43 0.20 Equation 4 -3.8 0.68 0.52 0.37 0.35 0.29 Equation 5 3.2 0.51 0.52 0.36 0.30 0.27 Equation 6 1.8 0.71 0.55 0.34 0.39 0.22
  • the sixth stage is the model matching and prediction stage.
  • the load data of the busbar on June 21, 2016 is selected as the test sample.
  • the weather data and date type information of the day are as follows:
  • the influencing factors corresponding to the 24th time of the day are respectively input into the random forest classification model, and the day to be tested belongs to the second type of load mode, so that the model 2 (regression equation 2) is selected to complete the load prediction, and the result is shown in FIG. .
  • the maximum relative error of the prediction result obtained by the method of the invention is 2.89%, the minimum relative error is 0.32%, and the average relative error is only 1.29%; while the maximum relative error of the traditional multiple regression prediction method is 3.06%, and the minimum relative error is 0.28. %, the average relative error is 1.38%. Therefore, the accuracy of the prediction method used in the present invention is significantly improved, and has broad application prospects.
  • DRAM dynamic RAM

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Medical Informatics (AREA)
  • Evolutionary Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种母线负荷预测方法,该方法包括:采用横向比较法修正历史负荷数据中的异常值(202),采用灰色关联投影法确定母线负荷的关键影响因素(203);采用改进的K-means聚类方法将特性相近的负荷曲线归为一类,得到若干典型负荷模式(204),构建随机森林分类模型,建立影响因素与聚类结果之间的映射关系(205);针对每类负荷模式,利用多元线性回归方法训练若干预测模型(206);通过随机森林分类模型确定待测日的类别,选择相匹配的回归模型实现负荷预测(207)。该方法引入数据挖掘方法分析母线负荷的变化规律并建立预测模型库,结合待测日的类别实现模型匹配,提高了短期母线负荷预测的精度和实时性,为电网规划和实时调度提供更加准确的决策支持。

Description

一种母线负荷预测方法 技术领域
本发明涉及电力系统工程技术领域,特别是指一种母线负荷预测方法。
背景技术
母线负荷预测是电力系统规划的重要组成部分,也是电力系统经济运行的基础,其预测结果能更好地实现分散式负荷管理,直接影响到电网后续安全校核的分析结果,对电网动态状态估计、无功优化、厂站局部控制和减少发电成本等方面具有重要意义。母线负荷作为变电站的下网负荷,其基数小,稳定性不强、变化趋势不明显,高频波动分量多等特点已成为提高预测精度的难点。而短期母线负荷预测一般指实时预测,不仅要求预测精度高,计算速度也要快,由于母线负荷自身的特点以及传统预测方法的固有缺陷,相对于系统负荷预测的研究比较少。
母线负荷预测一般采用回归分析法、神经网络和支持向量回归等方法。不同的预测方法从不同的角度提炼出历史负荷的信息,对样本数据规模和复杂度等方面的要求均不同。在处理考虑多影响因素的母线负荷预测问题时,回归分析法虽然实现简单,但对样本的平稳性要求较高,不宜直接处理波动性较强的母线负荷。神经网络和支持向量回归方法具有较强的非线性拟合能力,对样本的平稳性要求不高,但是当输入数据的维数过多,样本规模较大时,模型的训练容易陷入局部最优,且复杂度较高,从而降低预测精度和计算效率。而且随着大数据技术的兴起以及传感技术的快速发展,已形成电力大数据,母线负荷的内部特性更加复杂,不利于预测模型的建立。因此,现有的母线负荷的预测算法无法达到较为满意的预测效果。
发明内容
有鉴于此,本发明的目的在于提出一种母线负荷预测方法,能够分析母线负荷的变化规律,提高母线负荷的预测精度,同时满足实时性要求。
基于上述目的本发明提供的母线负荷预测方法,包括:
将待测母线的关键影响因素值作为特征向量,输入预先构建的分类模型中,得到待测日的类别,所述预先构建的分类模型为根据关键影响因素和对待测母线的历史负荷数据进行聚类分析的结果,建立的关键影响因素与母线负荷的映射关系;
根据待测日的类别从预先训练得到的多元回归预测模型中选取相应的多元回归预测模型,进行待测母线负荷的预测;所述预先训练得到的多元回归预测模型为根据待测母线的历史负荷数据的聚类分析结果,针对不同的典型负荷模式,选取对应的历史负荷数据和关键影响因素值作为样本数据训练得到的预测模型。
可选的,所述分类模型的构建方法为:
采用横向比较法筛选出待测母线历史负荷数据中的异常数据,将所述异常数据按照预设算法进行修正,得到修正负荷数据;
结合修正负荷数据以及与负荷相关的影响因素,采用灰色关联投影法计算得到影响因素与历史负荷数据的关联程度,由大到小选取预设数量的所述关联程度对应的影响因素作为关键影响因素;
按照预设的K-means聚类算法对待测母线的历史负荷数据进行聚类分析,得到不同的典型负荷模式;
根据所述关键影响因素和聚类分析的结果,采用随机森林算法构建分类模型,建立关键影响因素与各类母线负荷的映射关系。
可选的,所述采用横向比较法筛选出待测母线的历史负荷数据中的异常数据包括:
采用横向比较法对待测母线的历史负荷数据进行清洗,辨识出异常数据;其中,所述横向比较法采用相邻点负荷的变化率作为判断是否为异常数据的标准,所述负荷变化率计算公式如下:
Figure PCTCN2016108051-appb-000001
其中,α(d,t)为第d天t时刻的待测母线负荷变化率,L(d,t)为第d天t时刻的待测母线负荷值;采样点数据是否为异常数据的判断方法为:
Figure PCTCN2016108051-appb-000002
Figure PCTCN2016108051-appb-000003
其中,
Figure PCTCN2016108051-appb-000004
为第d天的前n天内t时刻负荷变化率的最大值,i表示与当前日d的间隔天数,L(d-i,t-1)为第d-i天t-1时刻的采样点数据,每天共有q+1个采样点;采用最近m天的数据对异常数据进行修正;
按时间顺序将当前时刻的负荷值与前一时刻的负荷值比较,当检测到异常数据时立即修正,并作为下一时刻负荷数据的比较值,完成负荷数据的修正。
可选的,所述将所述异常数据按照预设算法进行修正的方法为:
对筛选出的异常数据依据下列算法进行修正,具体算法公式为:
Figure PCTCN2016108051-appb-000005
Figure PCTCN2016108051-appb-000006
其中,λm为权值系数,用于表示第d-m天t时刻负荷对第d天t时刻负荷的影响程度;β为平滑系数,L(d-m,t)为第d-m天t时刻的待测母线负荷值。
可选的,所述采用灰色关联投影法计算得到影响因素与历史负荷数据的关联程度,由大到小选取预设数量的所述关联程度对应的影响因素作为关键影响因素,包括:
首先确定参考序列和比较序列,其中,所述参考序列为所述修正负荷数据:X0={x01,x02,...,x0m};比较序列为与负荷相关的影响因素:Xi={xi1,xi2,...,xim},i=1,2,...,n;得到参考序列和比较序列的序列矩阵:
Figure PCTCN2016108051-appb-000007
其中,m表示每一个序列有m个元素,n表示比较序列的个数;X0代表修正负荷数据对应的负荷序列或参考序列,Xi表示第i个比较序列;
对矩阵X进行标准化处理:
Figure PCTCN2016108051-appb-000008
其中,minj=1,2,...,mxij、maxj=1,2,...,mxij分别为xij所在行元素的最小值和最大值;eij为矩阵X中元素xij经标准化后的值,eij的值限定在0.1至1之间;
矩阵X经过标准化处理后为:
Figure PCTCN2016108051-appb-000009
其中,e0和ei,i=1,2,...,n为标准化后的标准化参考序列和标准化比较序列;
标准化比较序列ei与标准化参考序列e0的第k个元素的关联系数ξ0i(k)的计算公式为:
Figure PCTCN2016108051-appb-000010
其中,miniminj|eij-e0j|为两级最小差,maximaxj|eij-e0j|为两级最大差;ρ为分辨系数;
计算得到灰色关联系数判断矩阵F如下:
Figure PCTCN2016108051-appb-000011
其中,ξ00(j)为标准化参考序列e0与自身第j个元素的关联系数,值为1;
采用熵权法给矩阵E的列向量赋权值,第一行元素除外,计算公式如下:
Figure PCTCN2016108051-appb-000012
Figure PCTCN2016108051-appb-000013
其中,Ej表示第j列数据的信息熵,pij表示元素eij的值在该列元素中所占比重,wj表示第j列数据的权重,权值向量W=(w1,...,wj,...,wm).
得到加权灰色关联系数判断矩阵F':
Figure PCTCN2016108051-appb-000014
其中,wjξ0i(j)即序列ei与e0第j个元素的加权关联系数;
计算灰色关联投影值,计算公式如下:
Figure PCTCN2016108051-appb-000015
其中,Di代表比较序列Xi与待测母线负荷的关联投影值,Di在0到1之间;
由大到小选择投影值靠前的M个因素作为关键影响因素。
可选的,所述按照预设的K-means聚类算法对待测母线的历史负荷数据进行聚类分析的计算步骤为:
将所述修正负荷数据作为样本集合,设样本集合X={x1,...,xi,...,xm},负荷序列xi={xi0,...,xit,...,xin},初始质心集合
Figure PCTCN2016108051-appb-000016
随机选取一个样本作为第一质心,对剩余的每个样本分别计算与第一质心的欧氏距离:
Figure PCTCN2016108051-appb-000017
其中,m为负荷序列的个数,n表示每个负荷序列包括n+1个数据点,k代表质心个数;xit代表第i个负荷序列中t+1时刻的负荷值,
Figure PCTCN2016108051-appb-000018
代表第一质心向量中第t+1个元素值;
选取欧式距离最大的样本作为第二质心,然后继续计算剩余样本与第二质心的距离,然后继续选取欧式距离最大的样本作为第三质心,以此类推,直到k个质心都被确定下来;
分别计算每个样本与所有类型质心的欧氏距离,选定与当前样本的欧氏距离最小的质心作为当前样本所属的类别;
Figure PCTCN2016108051-appb-000019
其中,l为迭代次数,xi为第i个样本,
Figure PCTCN2016108051-appb-000020
为第l+1次迭代前第j类样本的质心向量;
当所有样本都归类后,更新每一类的质心;设第j类样本数量为mj,其样本集合为Sj,重新计算该类样本的质心向量为:
Figure PCTCN2016108051-appb-000021
其中,第t+1个分量
Figure PCTCN2016108051-appb-000022
计算公式为:
Figure PCTCN2016108051-appb-000023
判断是否达到预设的终止条件,若未终止则继续迭代计算,直到将待测母线的历史负荷归为k类典型负荷模式。
可选的,所述判断是否达到预设的终止条件为:在每一次迭代完成以后,按照下式计算各类质心更新前后的距离间隔
Figure PCTCN2016108051-appb-000024
表示第l次迭代后,第j类质心的距离间隔;
选出其中的最大值
Figure PCTCN2016108051-appb-000025
如果该最大值小于差异容忍度ε,则算法终止,否则重新将样本归类继续迭代更新。
可选的,所述采用随机森林算法构建分类模型的步骤包括:
对历史负荷数据样本集随机有放回抽取s个大小一样的训练样本集TSi,i=1,2,...,s;
随机抽取N个影响因素作为每个训练样本集的特征属性;其中,N<M;
分别训练s个样本集TSi,生成对应的决策树Treei;在整个森林生长的过程中,决策树Treei的特征属性保持不变;除叶子节点外,每个决策树采用CART算法中的基尼指数作为节点分裂依据:
Figure PCTCN2016108051-appb-000026
其中,d为当前节点,pi,d为目标类i出现的比例,k为目标类的个数;
对于每个节点,选择基尼指数最小的属性作为分裂属性a;当前节点d按分裂属性a划分的基尼指数为:
Gini(a,d)=pLGini(dL)+pRGini(dR),其中,dL和dR为节点d的左、右子节点,pL和pR为左右子节点的样本集在父节点中所占比例;
按照上述规则从上到下不断分裂,直到所有节点都被划分或被标记为叶子节点,即决策树生长完成,共有s棵,每棵树不需要剪枝操作,其叶子节点对应聚类分析中的某一聚类结果;将s棵决策树组合在起来,便得到随机森林分类模型;当输入一个测试样本,使用每一棵决策树对它分类,得到s个分类结果,选择出现比例最高的类作为该测试样本的类别。
可选的,所述多元回归预测模型采用多元线性回归法建立,回归参数通过最小二乘法确定。
可选的,回归方程为:
f(xi)=b0+b1xi1+b2xi2+…+bnxin
其中,b0为回归常数项,b1,b2,...,bn是回归系数,f(xi)为回归预测值;记每一类样本集中第i个训练样本为Xi={xi,yi},其中,xi={xi1,xi2,...,xin},i=1,2,...,m,m表示每一类负荷样本集中的样本个数;xi表示样本Xi的输入序列,xin为第n个影响因素值,yi为对应输出的母线负荷真实值;
记输入矩阵X、输出矩阵Y、系数矩阵B如下:
Figure PCTCN2016108051-appb-000027
采用最小二乘法确定回归方程的参数估计值:
Figure PCTCN2016108051-appb-000028
从上面所述可以看出,本发明提供的母线负荷预测方法通过灰色关联投影法确定负荷数据的关键影响因素,通过K-means聚类和随机森林分类方法,并结合回归分析法基于各类负荷序列的特征规律分别建立统计预测模型,挖掘母线负荷数据的内在规律,将变化复杂的历史负荷分解成几类典型负荷模式,有针对性的训练预测模型,从而根据待测日的类别选择相匹配的模型完成预测。所述母线负荷预测方法不仅能够提高母线负荷预测的精度,而且满足了预测的实时性要求,使得预测结果更为稳定、可靠。
附图说明
图1为本发明提供的母线负荷预测方法的一个实施例的流程图;
图2为本发明提供的母线负荷预测方法的另一个实施例的流程图;
图3为本发明提供的灰色关联投影法的一个实施例的流程图;
图4为本发明提供的预设K-means聚类算法的一个实施例的流程图;
图5为本发明提供的母线负荷预测方法中负荷数据的聚类结果图;
图6为本发明提供的随机森林算法的一个实施例的流程图;
图7为本发明提供的随机森林算法中的决策树生成缩略图;
图8为本发明提供的母线负荷预测方法中母线负荷预测的结果图。
具体实施方式
为使本发明的目的、技术方案和优点更加清楚明白,以下结合具体实施例,并参照附图,对本发明进一步详细说明。
需要说明的是,本发明实施例中所有使用“第一”和“第二”的表述均是为了区分两个相同名称非相同的实体或者非相同的参量,可见“第一”“第二”仅为了表述的方便,不应理解为对本发明实施例的限定,后续实施例对此不再一一说明。
为了克服现有技术中短期母线负荷预测存在的缺陷,本发明引入大数据技术中的数据挖掘方法聚类得到若干类母线负荷模式,并结合回归分析法基于各类负荷序列的特征规律分别建立统计预测模型,从而根据待测日的类别选择相应的模型完成预测,即实现模型匹配。与传统方法直接建立预测模型相比,该方法通过对历史负荷序列时域分解,更好地挖掘了母线负荷变化规律,有效提高预测精度;预测模型一旦训练完成,即得到预测模型库,可直接对母线一天内各时刻的负荷进行预测,不必每次预测前重新训练模型,满足短期预测的实时性要求。
参照图1所示,在本发明的一个实施例中,提供了一种母线负荷预测方法,包括:
步骤101,将待测母线当日的关键影响因素值作为特征向量,输入预先构建的分类模型中,得到待测日的类别;其中,所述预先构建的分类模型为根据关键影响因素和对待测母线的历史负荷数据进行聚类分析的结果,建立的关键影响因素与母线负荷的映射关系;
步骤102,根据待测日的类别从预先训练得到的多元回归预测模型中选取相应的多元回归预测模型,进行待测母线负荷的预测;其中,所述预先训练得到的多元回归预测模型为根据待测母线的历史负荷数据的聚类分析结果,针对不同的典型负荷模式,选取对应的历史负荷数据和关键影响因素值作为样本数据训练得到的预测模型。
这样,通过对待测日所属的模型类别先进行分类处理,再采用预测模型进行预测,能够使得母线负荷得到更为准确的预测,而且进一步提高了母线负荷预测的效率。尤其是对于短期母线负荷的预测。也即,本发明所述的母线负荷预测方法基于模型匹配实现了母线负荷的高效、准确的预测。
基于上述预先构建的分类模型为本发明的关键步骤之一,因此,在本发明一个可选的实施例中,还提供了一种分类模型的构建方法,包括:
采用横向比较法筛选出待测母线历史负荷数据中的异常数据,将所述异常数据按照预设算法进行修正,得到修正负荷数据;
结合修正负荷数据以及与负荷相关的影响因素,采用灰色关联投影法计算得到影响因素与历史负荷数据的关联程度,由大到小选取预设数量的所述关联程度对应的影响因素作为关键影响因素;
按照预设的K-means聚类算法对待测母线的历史负荷数据进行聚类分析,得到不同的典型负荷模式;
根据所述关键影响因素和聚类分析的结果,采用随机森林算法构建分类模型,建立关键影响因素与各类母线负荷的映射关系。
当然,上述构建方法中选定的算法只是作为本发明实施例一种可选的方式,还可以选用其他适合的算法进行模型的构建。同时也并不限定上述步骤之间的顺序,只需要使得逻辑关系符合正常的算法步骤即可。
参照图2所示,为本发明提供的母线负荷预测方法的另一个实施例的流程图。所述母线负荷预测方法包括:
步骤201,获取并存储待测母线的历史负荷数据以及与负荷相关的影响因素数据;其中,所述影响因素包括:气象信息、日类型等能够影响母线负荷的因素。
步骤202,采用横向比较法筛选出待测母线历史负荷数据中的异常数据,将所述异常数据按照预设算法进行修正,得到修正负荷数据;也即对母线负荷历史数据进行预处理,利用横向比较法对待测母线的历史负荷数据进行处理,辨识出负荷序列中的异常值,并给予修正。
步骤203,结合修正负荷数据以及与负荷相关的影响因素,采用灰色关联投影法计算得到影响因素与历史负荷数据的关联程度,由大到小选取预设数量的所述关联程度对应的影响因素作为关键影响因素;通过确定影响母线负荷的关键因素,使得预测模型有合理的输入维数。
步骤204,按照预设的K-means聚类算法对待测母线的历史负荷数据进行聚类分析,使初始聚类质心均匀分布在样本空间,并将变化特性相似的历史负荷数据归为一类,得到不同的典型负荷模式;其中,母线负荷聚类分析具体通过对待测母线的历史负荷数据以天为单位进行聚类分析,将变化特性相似的母线负荷归为一类,得到若干典型负荷模式。基于预设的K-means聚类算法将初始聚类质心均匀分布在样本空间,提高了算法的收敛速度和聚类准确性。
步骤205,根据所述关键影响因素和聚类分析的结果,采用随机森林算法构建分类模型,建立关键影响因素与各类母线负荷的映射关系;
步骤206,根据待测母线的历史负荷数据的聚类分析结果,针对不同的典型负荷模式,选取对应的历史负荷数据和关键影响因素值作为样本数据,训练得到不同的多元回归预测模型;
步骤207,将待测母线当日的关键影响因素值作为特征向量,输入所述分类模型中,得到待测日的类别,根据待测日的类别选取相应的多元回归预测模型,进行待测母线负荷的预测。
由上述实施例可知,本发明提供的母线负荷预测方法对待测母线历史负荷序列中的异 常值进行修正,通过灰色关联投影法确定母线负荷的关键影响因素;利用改进的K-means算法对修正后的历史数据聚类分析,并采用随机森林算法构建分类模型,建立聚类结果与关键影响因素的映射关系;针对每类负荷模式训练若干个多元回归预测模型;将待测日的影响因素值输入随机森林分类模型,得到待测日的类别,从而选择相匹配的回归模型完成预测。所述母线负荷预测方法不仅提高了母线负荷预测的精度,而且满足了预测的实时性要求,使得预测结果更为稳定、可靠。
在本发明一些可选的实施例中,所述采用横向比较法筛选出待测母线的历史负荷数据中的异常数据包括:采用横向比较法对待测母线的负荷历史数据进行清洗,辨识并修正异常值。横向比较法采用相邻点负荷的变化率作为评判异常数据的标准,负荷变化率计算公式如下:
Figure PCTCN2016108051-appb-000029
其中,α(d,t)是第d天t时刻的母线负荷变化率,L(d,t)为第d天t时刻的待测母线负荷值;
采样点数据是否为异常值的判断方法如下:
Figure PCTCN2016108051-appb-000030
Figure PCTCN2016108051-appb-000031
其中,
Figure PCTCN2016108051-appb-000032
为第d天前n天内t时刻负荷变化率的最大值,i表示与当前日d的间隔天数;L(d-i,t-1)为第d-i天t-1时刻的采样点数据,每天共有q+1个采样点。
对识别出的异常数据根据下式进行修正:
Figure PCTCN2016108051-appb-000033
Figure PCTCN2016108051-appb-000034
其中,λm为权值系数,用于表示第d-m天t时刻负荷对第d天t时刻负荷的影响程度,β为平滑系数,L(d-m,t)表示第d-m天t时刻的母线负荷值,即采用最近m天的数据作为修正数据。
整个过程按照时间顺序进行,只与前一时刻的负荷值进行比较,当检测到异常数据时立即修正,并作为下一时刻数据的比较值,完成所有历史负荷数据的修正。
在本发明一些可选的实施例中,参照图3所示,所述采用灰色关联投影法计算得到影 响因素与历史负荷数据的关联程度,由大到小选取预设数量的所述关联程度对应的影响因素作为关键影响因素,包括:采用灰色关联投影法计算每个影响因素与母线负荷的关联程度。将计算得到的灰色关联投影值进行排序,选择值较大的M个因素作为关键影响因素,具体计算步骤如下:
步骤301,首先确定参考序列和比较序列。参考序列为修正后的历史负荷序列X0={x01,x02,...,x0m},气象数据、日期类型等影响因素作为比较序列Xi={xi1,xi2,...,xim},i=1,2,...,n。因此可以得到序列矩阵如下:
Figure PCTCN2016108051-appb-000035
其中,m表示每一个序列有m个元素,n表示比较序列的个数。X0代表参考序列,即负荷序列,Xi表示第i个比较序列,即影响因素序列。
步骤302,对序列矩阵X按下式进行标准化处理:
Figure PCTCN2016108051-appb-000036
其中,minj=1,2,...,mxij代表xij所在行元素的最小值,maxj=1,2,...,mxij代表xij所在行元素的最大值。eij表示矩阵X中的元素xij经标准化后的值,eij的值一般限定在0.1至1之间。
矩阵X经过标准化处理后为:
Figure PCTCN2016108051-appb-000037
其中,e0和ei(i=1,2,...,n)为标准化后的参考序列和比较序列。
步骤303,计算关联系数。标准化比较序列ei与参考序列ei第k个元素的关联系数ξ0i(k),由下式决定:
Figure PCTCN2016108051-appb-000038
其中,miniminj|eij-e0j|为两级最小差,maximaxj|eij-e0j|为两级最大差,ρ为分辨系数,优选的,ρ取值为0.5。
由关联系数得到灰色关联系数判断矩阵F如下:
Figure PCTCN2016108051-appb-000039
其中,ξ00(j)表示标准化负荷序列e0与自身第j个元素的关联系数,所以值为1。
步骤304,采用熵权法给矩阵E的列向量(第一行元素除外)赋权值,其计算公式如下:
Figure PCTCN2016108051-appb-000040
Figure PCTCN2016108051-appb-000041
其中,Ej表示第j列数据的信息熵,pij表示元素eij的值在该列元素中所占比重,wj表示第j列数据的权重,权值向量W=(w1,...,wj,...,wm).
步骤305,由步骤304可得到加权灰色关联判断矩阵F'
Figure PCTCN2016108051-appb-000042
其中,wjξ0i(j)即序列ei与e0第j个元素的加权关联系数。
步骤306,计算灰色关联投影值,其计算公式如下:
Figure PCTCN2016108051-appb-000043
其中,Di代表比较序列Xi与母线负荷的关联投影值,取值在0到1之间,值越接近1,相关程度越大;值越接近0,相关程度越弱。与灰色关联度相比,灰色关联投影值可以更加综合地反应两个对象发展趋势的相似程度。
步骤307,最后选择投影值最大的M个影响因素作为关键影响因素。M值可以根据需要设定,且选取的过程可以先将所有投影值排序,由大到小选择投影值靠前的M个因素作 为关键影响因素,也可以不需要排序。
在本发明一些可选的实施例中,所述按照预设的K-means聚类算法对待测母线的历史负荷数据进行聚类分析的计算步骤包括:
将修正后的历史负荷数据作为样本集,采用改进的K-means算法用于母线负荷特性聚类。具体计算过程如下:
步骤401,初始聚类质心的选择。传统算法的初始聚类质心是随机选择的,而本发明提出的改进算法以距离为标准,将初始质心均匀地分布在样本空间,有效降低初始质心对聚类结果的影响。设样本集合X={x1,...,xi,...,xm},负荷序列xi={xi0,...,xit,...,xin},初始质心集合
Figure PCTCN2016108051-appb-000044
随机选取一个样本为第一质心。
步骤402,对剩余的每个样本,按下式分别计算与第一质心的欧氏距离d:
Figure PCTCN2016108051-appb-000045
其中,m表示样本集合中负荷序列的个数,n表示每个负荷序列包括n+1个数据点,k代表质心个数,即有k个负荷模式。xit代表第i个负荷序列中t+1时刻的负荷值,
Figure PCTCN2016108051-appb-000046
代表类1的初始质心向量
Figure PCTCN2016108051-appb-000047
中第t+1个元素值。
选择欧式距离最大的样本作为第2个质心,然后继续计算剩余样本与第2个质心的距离,选择距离最大的样本作为第3个质心,以此类推,直到k个质心都被确定下来。可选的,k为6。
步骤403,判断是否已经选出k个质心,若已将选出k个质心,则表示筛选操作已完成,可以进行下一步,若质心数量还未到达k个,则继续选取质心。
步骤404,将得到的k个质心分别与不同的类型的负荷模式相互对应,或者说建立对应关系,此时迭代的次数记为0。
步骤405,样本指派或样本确定。分别计算每个样本与所有质心的欧氏距离,筛选出与该样本的欧式距离最小的质心对应的负荷类型,将该样本指派给这负荷类型。计算公式如下:
Figure PCTCN2016108051-appb-000048
其中,l为迭代次数,xi为第i个样本,
Figure PCTCN2016108051-appb-000049
是第l+1次迭代前j类质心向量。
步骤406,更新质心。计算每一类的质心向量,并在所有样本都被指派后,开始更新每一类的质心。设第j类样本数量为mj,其样本集合为Sj,重新计算该类的质心向量为:
Figure PCTCN2016108051-appb-000050
其中,第t+1个分量
Figure PCTCN2016108051-appb-000051
计算如下,
Figure PCTCN2016108051-appb-000052
步骤407,判断是否达到终止条件。本发明中除选择设定最大迭代次数作为停止准则 外,还通过定义差异容忍度ε来判断算法是否可以终止。在每一次迭代完成以后,按照下式计算各类质心更新前后的距离间隔:
Figure PCTCN2016108051-appb-000053
表示第l次迭代后,第j类质心的距离间隔;
在每次指派更新后,会产生k种计算结果。选出其中的最大值
Figure PCTCN2016108051-appb-000054
如果该值小于预先定义的差异容忍度,则算法终止,否则转到步骤405继续迭代更新。
该过程将母线历史负荷归为k类负荷模式,每一种负荷模式下母线负荷的变化趋势更加明显、规律性更强,为构建模型和提高预测精度奠定了基础。
步骤408,得到k种聚类结果。
在本发明进一步的实施例中,所述采用随机森林算法构建分类模型的步骤还包括:
对历史样本集S随机有放回抽取s个与S一样的训练样本集TSi,i=1,2,...,s;
根据关键影响因素的个数M,随机抽取N(N<M)个影响因素作为每个训练样本集的特征属性;
分别训练s个样本集TSi,生成对应的决策树Treei。在整个森林生长的过程中,Treei的特征属性保持不变。除叶子节点外,每个决策树采用CART算法中的基尼指数(Gini index)作为节点分裂依据:
Figure PCTCN2016108051-appb-000055
其中,d为当前节点,pi,d为目标类i出现的比例,k为目标类的个数。基尼指数越接近0,则划分的纯度更高,效果越好。
对于每个节点,考虑各个特征属性对节点划分的所有情况,选择基尼指数最小的属性作为分裂属性a。当前节点d按分裂属性a划分的基尼指数为:
Gini(a,d)=pLGini(dL)+pRGini(dR),其中,dL和dR为节点d的左、右子节点,pL和pR为左右子节点的样本集在父节点中所占比例。
按照上述规则从上到下不断分裂,直到所有节点都被划分或被标记为叶子节点,即决策树生长完成,共有s棵,每棵树不需要剪枝操作,其叶子节点对应聚类分析中的某一聚类结果。当输入一个测试样本,便用每一棵决策树对它分类,得到s个分类结果,选择出现比例最高的类作为该测试样本的类别。
将s棵决策树组合在起来,便得到随机森林分类模型,从而建立影响因素与聚类结果之间的映射关系,实时高效地完成待测日类别的确定,为预测模型匹配奠定基础。
在本发明一些可选的实施例中,所述训练得到不同的多元回归预测模型的步骤还包括:
针对母线负荷的聚类结果,考虑到短期负荷预测的实时性,本发明选用多元线性回归模型进行预测。将每一类母线负荷数据和关键影响因素作为训练样本,构建若干个预测模型。预测模型采用多元线性回归法建立,参数通过最小二乘法确定。
具体的,记每类样本集中第i个训练样本为Xi={xi,yi},xi={xi1,xi2,...,xin},i=1,2,...,m,m表示每一类负荷样本集中的样本个数。xi表示样本Xi的输入序列,xin为第n个影响因素值,yi为对应输出的母线负荷真实值。
回归方程如下:
f(xi)=b0+b1xi1+b2xi2+…+bnxin
其中,b0为回归常数项,b1,b2,...,bn是回归系数,f(xi)为回归预测值。
记输入矩阵X、输出矩阵Y、系数矩阵B如下:
Figure PCTCN2016108051-appb-000056
采用最小二乘法确定回归方程的参数估计值:
Figure PCTCN2016108051-appb-000057
将待测日每一时刻的关键影响因素值输入到随机森林分类模型中,得到待测日的类别,从而选择相应的多元回归预测模型完成负荷预测。
本发明母线负荷预测方法具有以下效果:通过对历史母线负荷的聚类分析、统计模型的建立以及待测日的模型匹配,有效挖掘了母线负荷的变化规律,得到更加精确的预测结果;同时不必每次预测前训练模型,进一步满足母线负荷预测的实时性要求。
在本发明的另一个实施例中,以某电网地市公司一条110kV母线有功负荷作为预测实例。
第一阶段,为历史负荷数据预处理部分。选定某变电站110kV母线全天24时刻的负荷值作为为预测对象,并以预测日前一年的负荷数据为训练样本,数据格式如表1:
表1母线负荷数据表
Figure PCTCN2016108051-appb-000058
表1中每一行数据代表一个24维的负荷样本序列。由于采集系统的不稳定,历史负荷数据中可能存在若干异常值,即异常数据,会严重影响预测精度。
通过横向比较法对异常数据进行修正。采样周期设为1小时,每天共24个数据点,按照时间顺序依次计算相邻时刻负荷的变化率:
Figure PCTCN2016108051-appb-000059
将计算出的每一时刻采样点前7天内负荷变化率的最大值,作为该采样时刻数据是否为异常值的判据,判断公式如下:
Figure PCTCN2016108051-appb-000060
若负荷数据为异常值,则采用最近3天同一时刻的数据通过修正公式进行修正,修正公式如下:
L(d,t)=λ1L(d-1,t)+λ2L(d-2,t)+λ3L(d-3,t)
经过验证,当取平滑系数β=0.5,修正权值分别为λ1=0.5,λ2=0.26,λ3=0.24时,修正效果最好。
第二阶段,为关键影响因素的确定阶段。参照附图3,为灰色关联投影法的具体实现过程。以该地市公司所在城市的气象数据为基础,其影响因素选择为:最高温度(Tmax)、最低温度(Tmin),实时温度(RT),平均风速(AW)、相对湿度(RH)、平均降水量(AP)、日类型(DT)和季节类型(ST),具体参数如表2:
表2影响因素数据
Figure PCTCN2016108051-appb-000061
表2中,每种影响因素对应的列数据为比较序列Xi,各采样时刻的负荷数据为参考序列X0。日类型的赋值规则为{工作日,周末,节假日}={1,2,3},季节类型赋值为{夏季,冬季,春季,秋季}={4,3,2,1},其它影响因素数据都为实际测量值。
通过公式对比较序列和参考序列分别做标准化处理,将对应的数值限制在0.1至1之间,得到标准化序列矩阵E:
Figure PCTCN2016108051-appb-000062
其中,enm表示序列矩阵X中的元素xnm经标准化后的值。比较序列的个数n=8,m=8760为每个序列的元素个数,即每天24个采样点,共365天。
比较序列与参考序列各元素间的灰色关联系数由关联系数计算公式得到,从而得到灰色关联系数判断矩阵F。同时采用熵权法对矩阵E(除第一行元素)的列向量赋权值,得到 权值向量W=(w1,w2,...,w8760)。结合矩阵F,可得到加权灰色关联判断矩阵F'
Figure PCTCN2016108051-appb-000063
其中,ξ0i(j)为标准化比较序列ei与参考序列e0第j个元素的关联系数,wj为列向量j的权值,wjξ0i(j)即序列ei与e0第j个元素的加权关联系数。
参考序列Xi与比较序列X0的灰色关联投影值Di计算如下:
Figure PCTCN2016108051-appb-000064
通过上述计算,得到8种影响因素的灰色关联投影值如表3:
表3影响因素的灰色关联投影值
Tmax Tmin RT AW RH AP DT ST
0.87 0.64 0.72 0.28 0.66 0.32 0.55 0.43
平均风速和平均降雨量为弱相关性因素,日类型和季节类型为中等相关性因素。排序选取最高温度、实时温度、相对湿度、最低温度和日类型作为影响待测母线负荷走势的关键影响因素。
第三阶段,为待测母线的历史负荷聚类阶段。参照附图4,为改进的K-means算法具体实现过程。
确定初始聚类质心。历史负荷样本集合X共有365个样本序列,随机选取一个样本序列作为第一个初始质心向量
Figure PCTCN2016108051-appb-000065
分别计算剩余的每个样本与
Figure PCTCN2016108051-appb-000066
的欧氏距离d:
Figure PCTCN2016108051-appb-000067
其中,xit为第i个样本序列中t+1时刻的负荷值,
Figure PCTCN2016108051-appb-000068
为类1的初始质心向量中第t+1个元素值。
选择d最大的样本作为第二初始质心向量
Figure PCTCN2016108051-appb-000069
继续计算剩余样本与
Figure PCTCN2016108051-appb-000070
的距离,得到
Figure PCTCN2016108051-appb-000071
直到k个质心都被确定下来。本实施例中,k=6,将母线负荷序列归为6种典型负荷模式。
当6个初始质心确定以后,开始进行算法迭代,指派样本。计算每个样本与所有质心的欧氏距离,每个样本将被指派给离自己最近的那个质心所代表的类:
Figure PCTCN2016108051-appb-000072
其中,l为迭代次数,xi为第i个样本,
Figure PCTCN2016108051-appb-000073
是第l+1次迭代前类j的质心向量。
更新质心向量:每次指派样本后,需更新每一类负荷样本的质心。若类j样本数量为mj,其样本集合为Sj,则质心向量
Figure PCTCN2016108051-appb-000074
中的元素
Figure PCTCN2016108051-appb-000075
更新为:
Figure PCTCN2016108051-appb-000076
判断算法是否达到终止条件。除最大迭代次数之外,定义差异容忍度ε判断算法是否终止。在质心更新以后,计算每个类质心向量更新前和更新后的距离:
Figure PCTCN2016108051-appb-000077
选出其中的最大值
Figure PCTCN2016108051-appb-000078
若果该值小于差异容忍度ε=0.01,则算法终止,否则继续迭代更新。
上述聚类过程将母线历史负荷归为6种负荷模式,聚类结果如附图5所示。类1包含82天,主要集中在休息日,类2、类5主要包含112和89个工作日,这两类负荷值的差异主要是由于季节性因素造成的,类3包含25天,受极端天气所致,负荷远高于其他类;类4主要包含了所有的节假日,曲线形状明显和其它类不同;类6包含31天,负荷曲线波动较大。
第四阶段,为随机森林分类模型的构建阶段。参照附图6,为随机森林算法的具体实现过程。
对历史样本集S随机有放回抽取100个与S一样的训练样本集TSi,i=1,2,...,100。训练样本集TSi的数据如表4:
表4训练样本集
时间 Tmax RT RH Tmin DT 聚类结果
01/01 00:00 10 2 51 -2 3 类5
02/0509:00 5 3 64 -5 1 类2
针对每个训练样本集随机选取3个影响因素作为特征属性。
分别训练100个样本集TSi,生成对应的决策树Treei。在整个森林生长的过程中,Treei的特征属性保持不变。除叶子节点外,每个决策树采用CART算法中的基尼指数(Gini index)作为节点分裂依据:
Figure PCTCN2016108051-appb-000079
其中,d为当前节点,pi,d为目标类i出现的比例。基尼指数越接近0,则划分的纯度更高,效果越好。
对于每个节点,考虑3个特征属性对节点划分的所有情况,选择基尼指数最小的属性作为分裂属性a。当前节点d按属性a划分的基尼指数为:
Gini(a,d)=pLGini(dL)+pRGini(dR),其中,dL和dR为节点d的左、右子节点,pL和pR为左右子节点的样本集在父节点中所占比例。
按照上述规则从上到下不断分裂,直到所有节点都被划分或被标记为叶子节点,即决策树生长完成,共有100棵,每棵树不需要剪枝操作,其叶子节点对应某一聚类结果。当输入一个测试样本,便用每一棵决策树对它分类,得到100个分类结果,选择出现比例最高的类作为该测试样本的类别。
将100棵决策树组合在起来,便得到随机森林分类模型,从而建立影响因素与聚类结果之间的映射关系。附图7为随机森林分类模型中决策树生成的缩略图。
第五阶段,为预测模型的训练阶段。根据聚类结果,将每一负荷模式下的母线负荷数据和关键影响因素值整合为训练样本集,如表5:
表5训练样本集
Figure PCTCN2016108051-appb-000080
m为每一类样本集的样本个数。样本序列Xi={xi,yi},xi={xi1,xi2,...,xi5},i=1,2,...,m。其中,xi表示样本Xi的影响因素序列,yi为对应输出的母线负荷真实值。
每一类负荷模式需要根据自己的训练样本集构建预测模型,最终得到6个回归方程。多元回归方程由下式得到:
f(xi)=b0+b1xi1+b2xi2+…+b5xi5
其中,b0为回归常数项,b1,b2,...,b5是回归系数,f(xi)为回归预测值。每个回归方程的回归常数项和系数根据公式(22)计算得到,结果如表6:
表6回归系数
  b0 b1 b2 b3 b4 b5
方程1 4.9 0.64 0.51 0.36 0.32 0.23
方程2 -8.5 0.57 0.60 0.41 0.26 0.18
方程3 6.5 0.48 0.66 0.35 0.43 0.20
方程4 -3.8 0.68 0.52 0.37 0.35 0.29
方程5 3.2 0.51 0.52 0.36 0.30 0.27
方程6 1.8 0.71 0.55 0.34 0.39 0.22
第六阶段,为模型匹配与预测阶段。选定该母线2016年6月21日的负荷数据作为测试样本,当日的气象数据和日类型信息如下:
Figure PCTCN2016108051-appb-000081
Figure PCTCN2016108051-appb-000082
将该日24时刻对应的影响因素值分别输入到随机森林分类模型中,得到待测日属于第2类负荷模式,从而选择模型2(回归方程2)完成负荷预测,结果如附图8所示。采用本发明方法得到的预测结果最大相对误差为2.89%,最小相对误差为0.32%,而平均相对误差仅为1.29%;而传统多元回归预测法的最大相对误差为3.06%,最小相对误差为0.28%,平均相对误差为1.38%。因此,本发明中所采用的预测方法的准确度明显提升,具有广泛的应用前景。
所属领域的普通技术人员应当理解:以上任何实施例的讨论仅为示例性的,并非旨在暗示本公开的范围(包括权利要求)被限于这些例子;在本发明的思路下,以上实施例或者不同实施例中的技术特征之间也可以进行组合,步骤可以以任意顺序实现,并存在如上所述的本发明的不同方面的许多其它变化,为了简明它们没有在细节中提供。
另外,为简化说明和讨论,并且为了不会使本发明难以理解,在所提供的附图中可以示出或可以不示出与集成电路(IC)芯片和其它部件的公知的电源/接地连接。此外,可以以框图的形式示出装置,以便避免使本发明难以理解,并且这也考虑了以下事实,即关于这些框图装置的实施方式的细节是高度取决于将要实施本发明的平台的(即,这些细节应当完全处于本领域技术人员的理解范围内)。在阐述了具体细节(例如,电路)以描述本发明的示例性实施例的情况下,对本领域技术人员来说显而易见的是,可以在没有这些具体细节的情况下或者这些具体细节有变化的情况下实施本发明。因此,这些描述应被认为是说明性的而不是限制性的。
尽管已经结合了本发明的具体实施例对本发明进行了描述,但是根据前面的描述,这些实施例的很多替换、修改和变型对本领域普通技术人员来说将是显而易见的。例如,其它存储器架构(例如,动态RAM(DRAM))可以使用所讨论的实施例。
本发明的实施例旨在涵盖落入所附权利要求的宽泛范围之内的所有这样的替换、修改和变型。因此,凡在本发明的精神和原则之内,所做的任何省略、修改、等同替换、改进等,均应包含在本发明的保护范围之内。

Claims (10)

  1. 一种母线负荷预测方法,其特征在于,包括:
    将待测母线的关键影响因素值作为特征向量,输入预先构建的分类模型中,得到待测日的类别;所述预先构建的分类模型为根据关键影响因素和对待测母线的历史负荷数据进行聚类分析的结果,建立的关键影响因素与母线负荷的映射关系;
    根据待测日的类别从预先训练得到的多元回归预测模型中选取相应的多元回归预测模型,进行待测母线负荷的预测;所述预先训练得到的多元回归预测模型为根据待测母线的历史负荷数据的聚类分析结果,针对不同的典型负荷模式,选取对应的历史负荷数据和关键影响因素值作为样本数据训练得到的预测模型。
  2. 根据权利要求1所述的母线负荷预测方法,其特征在于,所述分类模型的构建方法为:
    采用横向比较法筛选出待测母线历史负荷数据中的异常数据,将所述异常数据按照预设算法进行修正,得到修正负荷数据;
    结合修正负荷数据以及与负荷相关的影响因素,采用灰色关联投影法计算得到影响因素与历史负荷数据的关联程度,由大到小选取预设数量的所述关联程度对应的影响因素作为关键影响因素;
    按照预设的K-means聚类算法对待测母线的历史负荷数据进行聚类分析,得到不同的典型负荷模式;
    根据所述关键影响因素和聚类分析的结果,采用随机森林算法构建分类模型,建立关键影响因素与各类母线负荷的映射关系。
  3. 根据权利要求2所述的母线负荷预测方法,其特征在于,所述采用横向比较法筛选出待测母线的历史负荷数据中的异常数据包括:
    采用横向比较法对待测母线的历史负荷数据进行清洗,辨识出异常数据;其中,所述横向比较法采用相邻点负荷的变化率作为判断是否为异常数据的标准,所述负荷变化率计算公式如下:
    Figure PCTCN2016108051-appb-100001
    其中,α(d,t)为第d天t时刻的待测母线负荷变化率,L(d,t)为第d天t时刻的待测母线负荷值;
    采样点数据是否为异常数据的判断方法为:
    Figure PCTCN2016108051-appb-100002
    Figure PCTCN2016108051-appb-100003
    其中,
    Figure PCTCN2016108051-appb-100004
    为第d天的前n天内t时刻负荷变化率的最大值,i表示与当前日d的间隔天数,L(d-i,t-1)为第d-i天t-1时刻的采样点数据,每天共有q+1个采样点;采用最近m天的数据对异常数据进行修正;
    按时间顺序将当前时刻的负荷值与前一时刻的负荷值比较,当检测到异常数据时立即修正,并作为下一时刻负荷数据的比较值,完成负荷数据的修正。
  4. 根据权利要求2所述的母线负荷预测方法,其特征在于,所述将所述异常数据按照预设算法进行修正的方法为:
    对筛选出的异常数据依据下列算法进行修正,具体算法公式为:
    Figure PCTCN2016108051-appb-100005
    Figure PCTCN2016108051-appb-100006
    其中,λm为权值系数,用于表示第d-m天t时刻负荷对第d天t时刻负荷的影响程度;β为平滑系数,L(d-m,t)为第d-m天t时刻的待测母线负荷值。
  5. 根据权利要求2所述的母线负荷预测方法,其特征在于,所述采用灰色关联投影法计算得到影响因素与历史负荷数据的关联程度,由大到小选取预设数量的所述关联程度对应的影响因素作为关键影响因素,包括:
    首先确定参考序列和比较序列,其中,所述参考序列为所述修正负荷数据:X0={x01,x02,...,x0m};比较序列为与负荷相关的影响因素:Xi={xi1,xi2,...,xim},i=1,2,...,n;得到参考序列和比较序列的序列矩阵:
    Figure PCTCN2016108051-appb-100007
    其中,m表示每一个序列有m个元素,n表示比较序列的个数;X0代表修正负荷数据对应的负荷序列或参考序列,Xi表示第i个比较序列;
    对矩阵X进行标准化处理:
    Figure PCTCN2016108051-appb-100008
    其中,minj=1,2,...,mxij、maxj=1,2,...,mxij分别为xij所在行元素的最小值和最大值;eij为矩阵X中元素xij经标准化后的值,eij的值限定在0.1至1之间;
    矩阵X经过标准化处理后为:
    Figure PCTCN2016108051-appb-100009
    其中,e0和ei,i=1,2,...,n为标准化后的标准化参考序列和标准化比较序列;
    标准化比较序列ei与标准化参考序列e0的第k个元素的关联系数ξ0i(k)的计算公式为:
    Figure PCTCN2016108051-appb-100010
    其中,miniminj|eij-e0j|为两级最小差,maximaxj|eij-e0j|为两级最大差;ρ为分辨系数;
    计算得到灰色关联系数判断矩阵F如下:
    Figure PCTCN2016108051-appb-100011
    其中,ξ00(j)为标准化参考序列e0与自身第j个元素的关联系数,值为1;
    采用熵权法给矩阵E的列向量赋权值,第一行元素除外,计算公式如下:
    Figure PCTCN2016108051-appb-100012
    Figure PCTCN2016108051-appb-100013
    其中,Ej表示第j列数据的信息熵,pij表示元素eij的值在该列元素中所占比重,wj表示第j列数据的权重,权值向量W=(w1,...,wj,...,wm).
    得到加权灰色关联系数判断矩阵F':
    Figure PCTCN2016108051-appb-100014
    其中,wjξ0i(j)即序列ei与e0第j个元素的加权关联系数;
    计算灰色关联投影值,计算公式如下:
    Figure PCTCN2016108051-appb-100015
    其中,Di代表比较序列Xi与待测母线负荷的关联投影值,Di在0到1之间,值越接近1,相关程度越大;值越接近0,相关程度越弱;
    由大到小选择投影值靠前的M个因素作为关键影响因素。
  6. 根据权利要求2所述的母线负荷预测方法,其特征在于,所述按照预设的K-means聚类算法对待测母线的历史负荷数据进行聚类分析的计算步骤为:
    将所述修正负荷数据作为样本集合,设样本集合X={x1,...,xi,...,xm},负荷序列xi={xi0,...,xit,...,xin},初始质心集合
    Figure PCTCN2016108051-appb-100016
    随机选取一个样本作为第一质心,对剩余的每个样本分别计算与第一质心的欧氏距离:
    Figure PCTCN2016108051-appb-100017
    其中,m表示样本集合中负荷序列的个数,n表示每个负荷序列包括n+1个数据点,k代表质心个数;xit代表第i个负荷序列中t+1时刻的负荷值,
    Figure PCTCN2016108051-appb-100018
    代表第一质心向量中第t+1个元素值;
    选取欧式距离最大的样本作为第二质心,然后继续计算剩余样本与第二质心的距离,然后选取欧式距离最大的样本作为第三质心,以此类推,直到k个质心都被确定下来;
    分别计算每个样本与所有类型质心的欧氏距离,选定与当前样本的欧氏距离最小的质心作为当前样本所属的类别;
    Figure PCTCN2016108051-appb-100019
    其中,l为迭代次数,xi为第i个样本,
    Figure PCTCN2016108051-appb-100020
    为第l+1次迭代前第j类样本的质心向量;
    当所有样本都归类后,更新每一类的质心;设第j类样本数量为mj,其样本集合为Sj,重新计算该类样本的质心向量为:
    Figure PCTCN2016108051-appb-100021
    其中,第t+1个分量
    Figure PCTCN2016108051-appb-100022
    计算公式为:
    Figure PCTCN2016108051-appb-100023
    判断是否达到预设的终止条件,若未终止则继续迭代计算,直到将待测母线的历史负荷归为k类典型负荷模式。
  7. 根据权利要求6所述的母线负荷预测方法,其特征在于,所述判断是否达到预设的终止条件为:
    在每一次迭代完成以后,按照下式计算各类质心更新前后的距离间隔:
    Figure PCTCN2016108051-appb-100024
    Figure PCTCN2016108051-appb-100025
    表示第l次迭代后,第j类质心的距离间隔;
    选出其中的最大值
    Figure PCTCN2016108051-appb-100026
    如果该最大值小于差异容忍度ε,则算法终止,否则重新将样本归类继续迭代更新。
  8. 根据权利要求2所述的母线负荷预测方法,其特征在于,所述采用随机森林算法构建分类模型的步骤包括:
    对历史负荷数据样本集随机有放回抽取s个大小一样的训练样本集TSi,i=1,2,...,s;
    随机抽取N个影响因素作为每个训练样本集的特征属性;其中,N<M;
    分别训练s个样本集TSi,生成对应的决策树Treei;在整个森林生长的过程中,决策树Treei的特征属性保持不变;除叶子节点外,每个决策树采用CART算法中的基尼指数作为节点分裂依据:
    Figure PCTCN2016108051-appb-100027
    其中,d为当前节点,pi,d为目标类i出现的比例,k为目标类的个数;
    对于每个节点,选择基尼指数最小的属性作为分裂属性a;当前节点d按分裂属性a划分的基尼指数为:
    Gini(a,d)=pLGini(dL)+pRGini(dR),其中,dL和dR为节点d的左、右子节点,pL和pR为左右子节点的样本集在父节点中所占比例;
    按照上述规则从上到下不断分裂,直到所有节点都被划分或被标记为叶子节点,即决策树生长完成,共有s棵,每棵树不需要剪枝操作,其叶子节点对应聚类分析中的某一聚类结果;将s棵决策树组合在起来,便得到随机森林分类模型;当输入一个测试样本,使用每一棵决策树对它分类,得到s个分类结果,选择出现比例最高的类作为该测试样本的类别。
  9. 根据权利要求1所述的母线负荷预测方法,其特征在于,所述多元回归预测模型采用多元线性回归法建立,回归参数通过最小二乘法确定。
  10. 根据权利要求9所述的母线负荷预测方法,其特征在于,回归方程为:
    f(xi)=b0+b1xi1+b2xi2+…+bnxin
    其中,b0为回归常数项,b1,b2,...,bn是回归系数,f(xi)为回归预测值;记每一类样本集中第i个训练样本为Xi={xi,yi},其中,xi={xi1,xi2,...,xin},i=1,2,...,m,m表示每一类负荷样本集中的样本个数;xi表示样本Xi的输入序列,xin为第n个影响因素值,yi为对应输出的母线负荷真实值;
    记输入矩阵X、输出矩阵Y、系数矩阵B如下:
    Figure PCTCN2016108051-appb-100028
    采用最小二乘法确定回归方程的参数估计值:
    Figure PCTCN2016108051-appb-100029
PCT/CN2016/108051 2016-09-09 2016-11-30 一种母线负荷预测方法 WO2018045642A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
AU2016325186A AU2016325186A1 (en) 2016-09-09 2016-11-30 Bus load forecasting method

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201610812339.0A CN106485262B (zh) 2016-09-09 2016-09-09 一种母线负荷预测方法
CN201610812339.0 2016-09-09

Publications (1)

Publication Number Publication Date
WO2018045642A1 true WO2018045642A1 (zh) 2018-03-15

Family

ID=58273572

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/108051 WO2018045642A1 (zh) 2016-09-09 2016-11-30 一种母线负荷预测方法

Country Status (3)

Country Link
CN (1) CN106485262B (zh)
AU (1) AU2016325186A1 (zh)
WO (1) WO2018045642A1 (zh)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109597967A (zh) * 2018-11-20 2019-04-09 江苏云上电力科技有限公司 一种电力系统配变电站负荷数据异常检测与修复方法
CN109636010A (zh) * 2018-11-23 2019-04-16 国网湖北省电力有限公司 基于相关因素矩阵的省级电网短期负荷预测方法及系统
CN109886465A (zh) * 2019-01-20 2019-06-14 东北电力大学 一种基于智能电表用户聚类分析的配电网负荷预测方法
CN110648248A (zh) * 2019-09-05 2020-01-03 广东电网有限责任公司 一种发电站的控制方法、装置及设备
CN110807508A (zh) * 2019-10-31 2020-02-18 国网辽宁省电力有限公司经济技术研究院 计及复杂气象影响的母线峰值负荷预测方法
CN110852628A (zh) * 2019-11-13 2020-02-28 国网江西省电力有限公司经济技术研究院 考虑发展模式影响的农村中长期负荷预测方法
CN111210059A (zh) * 2019-12-26 2020-05-29 国网北京市电力公司 母线日最高负荷的处理方法和装置
CN111505406A (zh) * 2020-03-16 2020-08-07 剑科云智(深圳)科技有限公司 一种配电柜、线材的监测方法
CN112487612A (zh) * 2020-11-12 2021-03-12 云南电网有限责任公司 基于分配因子和信息熵的母线负荷组合预测方法及系统
CN113779112A (zh) * 2021-09-27 2021-12-10 南京乐帆电气科技有限公司 基于空间信息和大数据映射算法的电能质量分析系统
CN113781254A (zh) * 2021-08-04 2021-12-10 广西电网有限责任公司 一种基于周用电量相关系数的异常用电检测算法
CN117977587A (zh) * 2024-04-02 2024-05-03 南京鼎研电力科技有限公司 基于深度神经网络的电力负荷预测系统及方法
CN117977587B (zh) * 2024-04-02 2024-06-07 南京鼎研电力科技有限公司 基于深度神经网络的电力负荷预测系统及方法

Families Citing this family (39)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107256415B (zh) * 2017-08-04 2020-12-11 国网经济技术研究院有限公司 一种电力系统运行方式场景的计算方法及计算系统
CN107800140B (zh) * 2017-10-18 2021-08-13 天津大学 一种考虑负荷特征的大用户供电接入决策方法
CN108334900B (zh) * 2018-01-29 2021-08-13 上海电气分布式能源科技有限公司 动力电池的分类模型的生成方法及系统、分类方法及系统
CN108539738B (zh) * 2018-05-10 2020-04-21 国网山东省电力公司电力科学研究院 一种基于梯度提升决策树的短期负荷预测方法
CN110766032A (zh) * 2018-07-27 2020-02-07 国网江西省电力有限公司九江供电分公司 基于分层递进策略的配电网数据聚类集成方法
CN109214458B (zh) * 2018-09-19 2021-08-13 合肥工业大学 一种基于历史数据的城市负荷量化方法
CN109345011A (zh) * 2018-09-19 2019-02-15 中冶赛迪重庆信息技术有限公司 一种基于深度回归森林的空调负荷预测方法及系统
CN109636035A (zh) * 2018-12-12 2019-04-16 北京天诚同创电气有限公司 负荷预测模型创建方法及装置、电力负荷预测方法及装置
CN111461378A (zh) * 2019-01-18 2020-07-28 电力规划总院有限公司 一种电网负荷预测方法和装置
CN109947079A (zh) * 2019-03-20 2019-06-28 阿里巴巴集团控股有限公司 基于边缘计算的区域异常检测方法和边缘计算设备
CN110175166B (zh) * 2019-04-18 2023-04-07 红云红河烟草(集团)有限责任公司 一种烟草卷包车间质量异常的确认方法
CN110570025A (zh) * 2019-08-20 2019-12-13 精硕科技(北京)股份有限公司 一种微信文章真实阅读率的预测方法、装置及设备
CN110489893B (zh) * 2019-08-23 2023-02-03 国网山东省电力公司莱芜供电公司 一种基于可变权值的母线负荷预测方法及系统
CN110598125A (zh) * 2019-08-26 2019-12-20 精硕科技(北京)股份有限公司 一种评估关键意见领袖投入的方法及装置、存储介质
CN110689195A (zh) * 2019-09-26 2020-01-14 云南电网有限责任公司电力科学研究院 一种电力日负荷预测方法
CN110826774B (zh) * 2019-10-18 2022-03-22 广东电网有限责任公司广州供电局 母线负荷预测方法、装置、计算机设备及存储介质
CN112862142A (zh) * 2019-11-28 2021-05-28 新奥数能科技有限公司 一种负荷、价格预测及修正方法
CN112882822B (zh) * 2019-11-29 2024-03-01 阿里巴巴集团控股有限公司 生成负载预测模型的方法、装置、设备和存储介质
CN111105098B (zh) * 2019-12-25 2023-11-03 国能信控互联技术有限公司 一种单体用户算法自匹配的负荷预测方法和系统
CN111476397B (zh) * 2020-03-05 2022-10-25 国电南瑞科技股份有限公司 应用于中长期电量安全校核的母线负荷预测方法及系统
CN111581883B (zh) * 2020-05-09 2022-09-23 国网上海市电力公司 一种在自动化装置上进行负荷计算与预测的方法
CN111626360B (zh) * 2020-05-28 2023-11-17 新奥新智科技有限公司 用于检测锅炉故障类型的方法、装置、设备和存储介质
CN112070307B (zh) * 2020-09-10 2023-01-20 云南电网有限责任公司电力科学研究院 一种区域内能源负荷量的预测方法及装置
CN112330009B (zh) * 2020-11-02 2023-09-01 广东电网有限责任公司 一种电网负荷预测方法
CN112289031A (zh) * 2020-11-03 2021-01-29 蚌埠学院 一种公交行驶过程中异常情况检测与报警的方法及装置
CN112365039B (zh) * 2020-11-03 2023-11-24 国网冀北电力有限公司计量中心 一种基于纵横相关性进行月度用电量预测的方法及系统
CN112327046B (zh) * 2020-11-09 2021-09-07 北华航天工业学院 基于模糊聚类和支持向量回归的非侵入式负荷监测方法
CN112446509B (zh) * 2020-11-10 2023-05-26 中国电子科技集团公司第三十八研究所 一种复杂电子装备预测维护方法
CN112529262A (zh) * 2020-11-27 2021-03-19 北京京能高安屯燃气热电有限责任公司 短期功率预测方法、装置、计算机设备和存储介质
CN112633738A (zh) * 2020-12-30 2021-04-09 安徽广志电气有限公司 一种配电控制柜终端负荷分配方法
CN113283774A (zh) * 2021-06-07 2021-08-20 润电能源科学技术有限公司 采暖供热机组深度调峰方法、装置、电子设备及存储介质
CN113393028B (zh) * 2021-06-10 2022-08-19 国网江苏省电力有限公司南通供电分公司 一种基于数据挖掘技术的负荷预测方法
CN113689053B (zh) * 2021-09-09 2024-03-29 国网安徽省电力有限公司电力科学研究院 一种基于随机森林的强对流天气架空线停电预测方法
CN113782187B (zh) * 2021-09-10 2023-06-27 深圳平安智慧医健科技有限公司 指标数据处理方法、相关设备及介质
CN113884807B (zh) * 2021-09-24 2023-10-20 广西电网有限责任公司电力科学研究院 基于随机森林和多层架构聚类的配电网故障预测方法
CN113888356A (zh) * 2021-10-19 2022-01-04 山东大学 一种建筑用能模式分析与判定方法及系统
CN114692386B (zh) * 2022-01-26 2024-02-06 天津大学 一种面向工业母线的运行状态异常监控方法
CN114819380A (zh) * 2022-05-12 2022-07-29 福州大学 基于模型融合的电网母线负荷预测方法
CN115085196B (zh) * 2022-08-19 2022-12-23 国网信息通信产业集团有限公司 电力负荷预测值确定方法、装置、设备和计算机可读介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103106329A (zh) * 2012-11-19 2013-05-15 华北电力大学 一种用于svr短期负荷预测的训练样本分组构造方法
CN104881706A (zh) * 2014-12-31 2015-09-02 天津弘源慧能科技有限公司 一种基于大数据技术的电力系统短期负荷预测方法
CN105303262A (zh) * 2015-11-12 2016-02-03 河海大学 一种基于核主成分分析和随机森林的短期负荷预测方法
US20160078457A1 (en) * 2011-07-19 2016-03-17 Ge Intelligent Platforms, Inc. System of Sequential Kernel Regression Modeling for Forecasting Financial Data
CN105469123A (zh) * 2015-12-30 2016-04-06 华东理工大学 一种基于k平面回归的缺失数据补全方法
CN105631483A (zh) * 2016-03-08 2016-06-01 国家电网公司 一种短期用电负荷预测方法及装置

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104123595B (zh) * 2014-07-22 2018-09-07 国家电网公司 一种配电网负荷预测方法及系统
CN104463445A (zh) * 2014-11-27 2015-03-25 国网浙江省电力公司嘉兴供电公司 一种基于经验日局部校正多点外推算法的负荷预测方法
CN105631532A (zh) * 2015-12-07 2016-06-01 江苏省电力公司检修分公司 一种基于模糊决策的神经网络模型电力系统负荷预测方法
CN105608512A (zh) * 2016-03-24 2016-05-25 东南大学 一种短期负荷预测方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160078457A1 (en) * 2011-07-19 2016-03-17 Ge Intelligent Platforms, Inc. System of Sequential Kernel Regression Modeling for Forecasting Financial Data
CN103106329A (zh) * 2012-11-19 2013-05-15 华北电力大学 一种用于svr短期负荷预测的训练样本分组构造方法
CN104881706A (zh) * 2014-12-31 2015-09-02 天津弘源慧能科技有限公司 一种基于大数据技术的电力系统短期负荷预测方法
CN105303262A (zh) * 2015-11-12 2016-02-03 河海大学 一种基于核主成分分析和随机森林的短期负荷预测方法
CN105469123A (zh) * 2015-12-30 2016-04-06 华东理工大学 一种基于k平面回归的缺失数据补全方法
CN105631483A (zh) * 2016-03-08 2016-06-01 国家电网公司 一种短期用电负荷预测方法及装置

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109597967A (zh) * 2018-11-20 2019-04-09 江苏云上电力科技有限公司 一种电力系统配变电站负荷数据异常检测与修复方法
CN109597967B (zh) * 2018-11-20 2023-08-25 江苏云上电力科技有限公司 一种电力系统配变电站负荷数据异常检测与修复方法
CN109636010A (zh) * 2018-11-23 2019-04-16 国网湖北省电力有限公司 基于相关因素矩阵的省级电网短期负荷预测方法及系统
CN109636010B (zh) * 2018-11-23 2022-12-02 国网湖北省电力有限公司 基于相关因素矩阵的省级电网短期负荷预测方法及系统
CN109886465B (zh) * 2019-01-20 2022-03-18 东北电力大学 一种基于智能电表用户聚类分析的配电网负荷预测方法
CN109886465A (zh) * 2019-01-20 2019-06-14 东北电力大学 一种基于智能电表用户聚类分析的配电网负荷预测方法
CN110648248A (zh) * 2019-09-05 2020-01-03 广东电网有限责任公司 一种发电站的控制方法、装置及设备
CN110648248B (zh) * 2019-09-05 2023-04-07 广东电网有限责任公司 一种发电站的控制方法、装置及设备
CN110807508A (zh) * 2019-10-31 2020-02-18 国网辽宁省电力有限公司经济技术研究院 计及复杂气象影响的母线峰值负荷预测方法
CN110807508B (zh) * 2019-10-31 2023-06-09 国网辽宁省电力有限公司经济技术研究院 计及复杂气象影响的母线峰值负荷预测方法
CN110852628A (zh) * 2019-11-13 2020-02-28 国网江西省电力有限公司经济技术研究院 考虑发展模式影响的农村中长期负荷预测方法
CN110852628B (zh) * 2019-11-13 2023-09-12 国网江西省电力有限公司经济技术研究院 考虑发展模式影响的农村中长期负荷预测方法
CN111210059A (zh) * 2019-12-26 2020-05-29 国网北京市电力公司 母线日最高负荷的处理方法和装置
CN111505406A (zh) * 2020-03-16 2020-08-07 剑科云智(深圳)科技有限公司 一种配电柜、线材的监测方法
CN112487612B (zh) * 2020-11-12 2022-08-30 云南电网有限责任公司 基于分配因子和信息熵的母线负荷组合预测方法及系统
CN112487612A (zh) * 2020-11-12 2021-03-12 云南电网有限责任公司 基于分配因子和信息熵的母线负荷组合预测方法及系统
CN113781254A (zh) * 2021-08-04 2021-12-10 广西电网有限责任公司 一种基于周用电量相关系数的异常用电检测算法
CN113779112A (zh) * 2021-09-27 2021-12-10 南京乐帆电气科技有限公司 基于空间信息和大数据映射算法的电能质量分析系统
CN117977587A (zh) * 2024-04-02 2024-05-03 南京鼎研电力科技有限公司 基于深度神经网络的电力负荷预测系统及方法
CN117977587B (zh) * 2024-04-02 2024-06-07 南京鼎研电力科技有限公司 基于深度神经网络的电力负荷预测系统及方法

Also Published As

Publication number Publication date
AU2016325186A1 (en) 2018-03-29
CN106485262B (zh) 2020-02-07
CN106485262A (zh) 2017-03-08

Similar Documents

Publication Publication Date Title
WO2018045642A1 (zh) 一种母线负荷预测方法
CN111199016B (zh) 一种基于DTW的改进K-means的日负荷曲线聚类方法
CN110991786B (zh) 基于相似日负荷曲线的10kV静态负荷模型参数辨识方法
CN109002915B (zh) 基于Kmeans-GRA-Elman模型的光伏电站短期功率预测方法
CN101414366B (zh) 基于改进极端学习机方法的电力系统短期负荷预测方法
CN101383023B (zh) 基于样本动态组织与温度补偿的神经网络短期电力负荷预测
CN109376772B (zh) 一种基于神经网络模型的电力负荷组合预测方法
CN111401599B (zh) 一种基于相似性搜索和lstm神经网络的水位预测方法
CN115270965B (zh) 一种配电网线路故障预测方法和装置
CN111369070A (zh) 一种基于包络线聚类的多模融合光伏功率预测方法
CN103049651A (zh) 一种用于电力负荷聚类的方法及装置
CN109492748B (zh) 一种基于卷积神经网络的电力系统的中长期负荷预测模型建立方法
CN111160750A (zh) 一种基于关联规则挖掘的配网分析和投资决策方法
CN110188221B (zh) 一种基于形状距离的负荷曲线层次聚类方法
CN113468817B (zh) 一种基于igoa优化elm的超短期风电功率预测方法
CN111369045A (zh) 一种短期光伏发电功率预测的方法
CN111461921B (zh) 一种基于机器学习的负荷建模典型用户数据库更新方法
CN112418476A (zh) 一种超短期电力负荷预测方法
CN111882114B (zh) 一种短时交通流量预测模型构建方法及预测方法
CN111539657B (zh) 结合用户日用电量曲线的典型用电行业负荷特性分类与综合方法
CN115759389A (zh) 基于天气类型的相似日组合策略的日前光伏功率预测方法
CN114936694A (zh) 一种基于双集成模型的光伏功率预测方法
CN111027841A (zh) 一种基于梯度提升决策树的低压台区线损计算方法
CN110163437B (zh) 基于DPK-means的日前光伏发电功率预测方法
CN106960260B (zh) 一种便于电力调度的风电功率预测系统

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2016325186

Country of ref document: AU

Date of ref document: 20161130

Kind code of ref document: A

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 16915576

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 16915576

Country of ref document: EP

Kind code of ref document: A1