CN105117810A - Residential electricity consumption mid-term load prediction method under multistep electricity price mechanism - Google Patents

Abstract

The invention relates to a residential electricity consumption mid-term load prediction method under a multistep electricity price mechanism. At first, the residential electricity consumption data is acquired, the attributive characters of residential electricity consumption behaviors under the multistep electricity price mechanism are extracted, different electricity consumption behavior characters of the residents under the multistep electricity price mechanism can be identified through cluster analysis, and the users having same or similar electricity consumption behavior characters are clustered into the same user category; corresponding load prediction models are established for the user categories, and the predication is carried out; at the end, the predication results of the user categories are summarized. The multistep electricity price related indexes are innovatively introduced into a clustering model, and the mid-term load prediction can be better carried out by means of the accurate and comprehensive data provided by a smart electric meter. Compared with a conventional manner that the data acquisition is carried out every 15 minutes, the method provided performs the mid-term prediction with fewer data sampling points (data per day) under the premise of guaranteeing the precision.

Description

Residential electricity consumption load predicting method in mid-term under a kind of step price mechanism

Technical field

The present invention relates to load estimation and Data Mining, more particularly, relate to residential electricity consumption load predicting method in mid-term under a kind of step price mechanism.

Background technology

Resident's step price, since enforcement, has achieved preliminary effect, and has enhanced the awareness of saving energy of resident to a certain extent, changed the consumption habit that some are bad.Under the power price system of traditional single low price, electricity consumption behavior (but not power consumption) difference between the customer group of different characteristic is also not obvious.

But under new step price system, (as income, family structure, habits and customs etc.) customer group of different characteristic will produce different responses to step price, and the electricity consumption behavior difference caused by this will progressively highlight.This also makes the load estimation under step price become more complicated.Accurate load estimation can not only ensure that network system is run with security and stability, can also reduce power operation cost simultaneously, improve the economic and social benefits.

Load estimation can be divided into short-term, mid-term and long-term according to time domain.Short-term refers generally to the prediction of coming few hours, a day to several days, then refers to the prediction of following several weeks, several months mid-term, and long-term forecasting is then to prediction over the next several years, even for more time.Compared to short-term forecasting, medium-term and long-term load time span is longer, and the basic data amount of needs is comparatively large, and easily by the interference of many factors, predicated error is accumulated and becomes unreliable.

The scientific theory basis of step price distinguishes different characteristic customer group by the method for the market segments, mines massively and use different Price Mechanisms, to improve Allocation Efficiency for different user.China is still in the elementary step of comprehensively carrying out step price, also more rare to the research of electricity consumption customer-action analysis aspect under China's step price, theory and practice all exists larger research blank.

Since the eighties mid-term, Chinese scholars carries out the electrical load forecasting research in a large number based on various electrical load forecast model and method.But the overwhelming majority is single load estimation model, mainly comprises the models such as regretional analysis, time series, neural network and support vector machine.Single load estimation model, by the key factor of analyzing influence region power consumption, catches the correlationship between variable, and constructs model with this and predict.But the power consumption of user just simply adds up by nearly all single load estimation model, with region electricity consumption total amount for target sets up the average behavior model in this region, and have ignored the characteristic of dissimilar user power utilization behavior.

In existing research, subscriber segmentation method is all generally based on the unitary variant such as family income, power consumption, customer group is simply divided into the customer group of high, medium and low income.Meanwhile, the number determining customer group is in advance needed.Such as according to family's year per capita income by residential households people for being divided into four classes; With the monthly power consumption of user for sole indicator, according to the dwelling density in monthly power consumption neighborhood, resident is tentatively divided into basic, normal, high income three class user.But user's monthly electricity consumption situation is not unalterable, and changes along with the change in temperature, season often.The factor numerous and complicated affecting electricity consumption has been pointed out in many research in the recent period, more the variation of horn of plenty needs to consider in subscriber segmentation, as segmented power industry client by segmentation variablees such as annual electricity consumption total amount, average electricity price, electricity consumption rate of growth, the coefficient of variation, load factor, paying rates.But, also step price correlated variables is not joined in subscriber segmentation in existing research.On the other hand, the research both at home and abroad about load forecast is very many, studies in the past and always the electricity consumption total amount of certain district's intra domain user is predicted as target.The average behavior model of what traditional Prediction of Total was set up is user, this class model cannot disclose the different behaviors of each customer group, have ignored the otherness of dissimilar user power utilization behavior simultaneously.Therefore, efficiently and accurately class of subscriber and catch all types of user electricity consumption Behavior law be under step price mechanism intelligent grid planning be badly in need of improve two importances.

Traditional load estimation is all carry out load estimation according to the historical data of total amount, in recent years, along with popularizing of intelligent electric meter, existing resident is in real time by the real-time electricity consumption data of abundanter and meticulous resident that extraction system can catch under step price quickly and easily, and this is for identifying that dissimilar user's provides strong Data support with electrical feature.

At present, common combination load predictive mode sets up multiple different forecast model to same inputoutput data collection, then will predict the outcome and combine, or be weighted on average by suitable weight, or adopt more complicated nonlinear combination model, finally select the built-up pattern that degree of fitting is best or standard deviation is minimum.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, providing a kind of accurately centering permanent load can carry out residential electricity consumption load predicting method in mid-term under the step price mechanism predicted.

Technical scheme of the present invention is as follows:

Residential electricity consumption load predicting method in mid-term under a kind of step price mechanism, first, gather residential electricity consumption data, extract the attributive character of residential electricity consumption behavior under step price mechanism, identifying the different electricity consumption behavioural characteristic of resident under step price mechanism by cluster analysis, is same class of subscriber by the user Gui Ju with same or analogous electricity consumption behavioural characteristic;

Then, set up corresponding load estimation model respectively for each class of subscriber, and predict;

Finally, predicting the outcome of each class of subscriber is gathered.

As preferably, first residential electricity consumption data are divided into several classes by different electricity consumption behavioural characteristics, obtain one group several there is the data set of different input and output, then corresponding load estimation model is set up to each data set.

As preferably, after gathering residential electricity consumption data, proceed as follows:

1) data prediction: obtain each user and work as daily power consumption;

2) missing values process: if lack certain daily power consumption, then by calculating before electricity consumption missing time section the difference of accumulative electricity one day after, and be averaged according to the number of days of disappearance, as certain daily power consumption of disappearance;

3) outlier processing: the daily power consumption of working as of the metrics-thresholds scope exceeding setting is filtered.

As preferably, when predicting, first to identify the class of subscriber that user to be predicted returns, then select corresponding load estimation model to carry out load estimation, finally, gather single predicting the outcome, obtain final macro-forecast result.

As preferably, the attributive character extracting residential electricity consumption behavior comprises cluster Attributions selection, prediction input variable is extracted; Cluster attribute comprises the average daily power consumption of each user, the second ladder ratio, the 3rd ladder ratio and sensitive; Prediction input variable comprises the power consumption of seven days, the temperature on the same day in the past.

As preferably, average daily power consumption=total electricity consumption/total number of days of sampling;

Second ladder ratio=the arrive second ladder month number/total moon number;

3rd ladder ratio=arrival the 3rd ladder moon number/total moon number;

The average power consumption of sensitive=day with high temperature/average daily power consumption.

As preferably, cluster analysis is realized by fuzzy c means means clustering algorithm, and each attributive character of user is under the jurisdiction of one or more class of subscriber, represents that it belongs to the degree of different user classification with degree of membership.

As preferably, load estimation model, based on Self-organized Fuzzy Neural Network model, comprises input layer, ellipsoidal function layer, normalization layer, weighted mean layer, output layer.

As preferably, the learning process of Self-organized Fuzzy Neural Network comprises parameter learning, Structure learning;

Parameter learning makes network Fast Convergent by online Recursive Least Squares;

Structure learning, by automatically increasing, revising or deleting the neuronic self-organization in ellipsoidal function layer, is searched and selects scale of neural network.

As preferably, Structure learning comprises following operation:

1) neuron is increased;

2) neuron is pruned;

3) membership function in ellipsoidal function layer and fuzzy rule is merged.

Beneficial effect of the present invention is as follows:

Method of the present invention, from model mechanism, is different from traditional total amount load estimation model; Innovatively index relevant for step price is introduced in Clustering Model; What utilize intelligent electric meter to provide is more accurate, and more fully data, carry out load estimation in mid-term better.Meanwhile, compare traditional every 15 minute data collections, under the prerequisite ensureing accuracy requirement, this method utilizes less data acquisition sampling point (data every day) to carry out medium-term forecast.But accurate medium-term and long-term load, contributing to providing a series of decision support for carrying out intelligent distribution network planning construction scientifically and rationally further, is the important module realizing intelligent grid.

Method of the present invention with cluster analysis and load estimation algorithm for core, a kind of classification load estimation model is proposed, combination combines the methods such as fuzzy C-means clustering (FCM) and Self-organized Fuzzy Neural Network (SOFNN), the different characteristic of user power utilization behavior under step price mechanism can not only be caught, and overall load estimation precision in mid-term is increased.Accurately mid-term load estimation.

Classification load estimation model proposed by the invention and traditional combined prediction difference, be that in (1) classification load estimation model, each submodel is the load of prediction one class, but the input and output predicted are all different; (2) predicting the outcome of each submodel is just gathered, not the modeling process of the weighting of combination forecasting neutral line or nonlinear combination; (3) load estimation model of classifying can obtain two classes easily and export, and except traditional Prediction of Total result, can also obtain predicting the outcome of each electricity consumption classification, but combination forecasting can only obtain Prediction of Total result.

Accompanying drawing explanation

Fig. 1 is principle flow chart of the present invention (comprising the framework of load estimation model);

Fig. 2 is the basic framework of Self-organized Fuzzy Neural Network;

Fig. 3 is different classes of SSE and MIA index test Comparative result figure;

Fig. 4 is the electricity consumption of resident curve family after cluster;

Fig. 5 is that test set predicts the outcome comparison diagram.

Embodiment

Below in conjunction with drawings and Examples, the present invention is described in further detail.

Residential electricity consumption load predicting method in mid-term under a kind of step price mechanism of the present invention, as shown in Figure 1, first, gather residential electricity consumption data, extract the attributive character of residential electricity consumption behavior under step price mechanism, identifying the different electricity consumption behavioural characteristic of resident under step price mechanism by cluster analysis, is same class of subscriber by the user Gui Ju with same or analogous electricity consumption behavioural characteristic;

Then, set up corresponding load estimation model respectively for each class of subscriber, and predict;

Finally, predicting the outcome of each class of subscriber is gathered.

The core of method of the present invention is a kind of classification load estimation model, first residential electricity consumption data are divided into several classes by different electricity consumption behavioural characteristics, obtain one group several there is the data set of different input and output, then corresponding load estimation model is set up to each data set.Classification load estimation model is by first classifying to all information contained of single load estimation model, and different electricity consumption behavioural characteristics adopts different forecast models, and this contributes to providing and predicts the outcome more accurately.Such as, there is the customer group that two different, be respectively the customer group low with price sensitivity that price sensitivity is high, classification prediction can know the accurate behavior of each customer group by let us, the averaging model of traditional total amount then can cause the customer group prediction to price sensitivity is high too high, and the low customer group prediction of price sensitivity is too low.

After gathering residential electricity consumption data, proceed as follows, complete data preparation stage:

1) data prediction: obtain each user and work as daily power consumption;

2) missing values process: if lack certain daily power consumption, then by calculating before electricity consumption missing time section the difference of accumulative electricity one day after, and be averaged according to the number of days of disappearance, as certain daily power consumption of disappearance;

3) outlier processing: the daily power consumption of working as of the metrics-thresholds scope exceeding setting is filtered.

Particularly, after completing data preparation stage, utilize clustering algorithm to catch and identify the electricity consumption behavioural characteristic of different user within the analysis phase, same or analogous for electricity consumption behavioural characteristic user is polymerized to a class.Then classify to residential electricity consumption data, namely the input and output of each data class are different.Then, optimal load estimation model is set up respectively to dissimilar user's (i.e. different pieces of information collection).

When predicting, first to identify the class of subscriber that user to be predicted returns, then select corresponding load estimation model to carry out load estimation, finally, gather single predicting the outcome, obtain final macro-forecast result.

The attributive character extracting residential electricity consumption behavior mainly comprises cluster Attributions selection, prediction input variable is extracted.

Through data analysis and preliminary experiment, the present invention is extracted the average daily power consumption of each user, the second ladder ratio, the 3rd ladder ratio and sensitive four groups of cluster attributes, reflects the load variations rule in resident's certain hour section.

Average daily power consumption depends primarily on the number of all kinds of electric equipments that family has, and can infer the income level of resident whereby.And the close residential electricity consumption behavior of income level is often more similar.

Second ladder ratio, the 3rd ladder ratio then can reflect the user's interior undulatory property of electricity consumption and reflection to step price mechanism over the past several months, the electricity consumption rule that under catching step price mechanism, each user is long-term.Such as, the user comparatively responsive to price, when entering next ladder soon, can reduce power consumption consciously, to avoid entering next ladder, thus reducing total electricity price.

In addition, meteorologic factor, especially temperature, often have an impact to residential electricity consumption change.Especially Summer High Temperature, the household electrical appliance frequencies of utilization such as air-conditioning are higher, and daily power consumption often significantly improves.Sensitive index reflection user is in the fluctuation situation of hot weather power consumption.

The computing method of above-mentioned four groups of cluster attributes are:

Average daily power consumption=total electricity consumption/total number of days of sampling;

Second ladder ratio=the arrive second ladder month number/total moon number;

3rd ladder ratio=arrival the 3rd ladder moon number/total moon number;

The average power consumption of sensitive=day with high temperature/average daily power consumption; Wherein, the day with high temperature that the present embodiment defines refers to that medial temperature is more than or equal to the date of 25 DEG C.

Meanwhile, the input variable of all load estimation models is mainly extracted over the power consumption of seven days and the temperature on the same day and amounts to eight input attributes.Wherein, applicating history load data is useful to prediction, because rolling forecast mode can be used to carry out.And if temperature is unknown and need prediction, the medial temperature of can use data of weather forecast or this area same day is in the past few years estimated.

The classification load estimation model that the present invention proposes has versatility and the good feature of compatibility, be applicable to different clustering method and load estimation model independent assortment under this framework, comprise conventional clustering method (K average, self-organizing feature map neural network), load estimation model (return, time series, support vector machine).In addition, different class of subscribers can adopt diverse load estimation model to carry out load estimation, contributes to adopting optimal load estimation model to each class of subscriber, will greatly improve dirigibility and the precision of prediction of classification load estimation model.

In the present embodiment, use Fuzzy C-Means Cluster Algorithm (FCM) and the method for Self-organized Fuzzy Neural Network (SOFNN) models coupling, carry out classification load estimation.

Compared to traditional K-means clustering algorithm, FCM adds fuzzy concept, makes each input vector (attributive character) no longer only be under the jurisdiction of some specific clusters, but represents that it belongs to the degree of different cluster with its degree of membership.Namely each attributive character of user is under the jurisdiction of one or more class of subscriber, represents that it belongs to the degree of different user classification with degree of membership.

In addition, the advantage of SOFNN is: the first, is simple and easy to use, even if user is familiar with dark to fuzzy system and neural network, SOFNN also can the structure of Confirming model automatically, the parameter of model of cognition; The second, precision of prediction is higher.

The basic thought of FCM obtains the degree of membership of each sample point for all cluster centres by continuing to optimize objective function, and then determine the generic of sample point, finally reaches the object automatically to sample data cluster.

Suppose that sample set is Z={z ₁, z ₂..., z _n, N is total sample number.Be divided into C fuzzy clustering group, and obtain cluster centre and integrate as V={v ₁, v ₂..., v _c, according to principle of least square method, adopt following optimization object function to carry out dividing data:

$J\left(U,V\right)=\underset{c=1}{\overset{C}{\Σ}}\underset{n=1}{\overset{N}{\Σ}}{\left(u_{cn}\right)}^m\left|\right|z_n-v_c|{|}^2,n=1,2,\mathrm{...}N,c=1.2,\mathrm{...}C;$

Wherein, m is Fuzzy tuning parameter, u _cnthe degree of membership of the n-th sample c class, and 0≤u _cn≤ 1, U=[U _cn] be the matrix that C × N ties up.

In addition, before fuzzy clustering, the characteristic attribute extracted need need be normalized, the property value by these is mapped between [0.1], to remove the impact of different magnitude on user power utilization measure feature.Usual employing Min-max method is normalized data set, and disposal route is as follows:

${z_n}^{\′}=\frac{z_n-z_{n_{\min }}}{z_{n_{\max }}-z_{n_{\min }}};$

Z in formula _n' for adopting the n-th sample data after the normalization of Min-max method, with be respectively maximal value and the minimum value of data sequence.

Load estimation model is based on Self-organized Fuzzy Neural Network model, and shown in Fig. 2, SOFNN model comprises input layer, ellipsoidal function (EBF) layer, normalization layer, weighted mean layer and output layer and forms.Wherein,

(1) the neuron i=1 in input layer, 2 ..., r represents input variable x _i;

(2) the neuron j=1 in EBF layer, 2 ..., u represents the prerequisite of a fuzzy rule, and the value of wherein all subordinate functions is multiplied as Output rusults Φ by each neuron _j, specific algorithm is as follows:

${\mathrm{\Φ}}_j=\exp \[-\underset{i=1}{\overset{r}{\Σ}}\frac{{\left(x_i-c_{ij}\right)}^2}{2{\mathrm{\δ}}_{ij}^2}\];$

In formula, c _ijrepresent the center of subordinate function, δ _ijrepresent the width of subordinate function;

(3) neuron number in normalization layer is general identical with EBF layer, the Output rusults Ψ of its correspondence _jfor:

${\mathrm{\Ψ}}_j=\frac{{\mathrm{\Φ}}_j}{{\mathrm{\Σ}}_{k=1}^u{\mathrm{\Φ}}_k}=\frac{\exp \[-{\Σ}_{i=1}^r\frac{{\left(x_i-c_{ij}\right)}^2}{2{\mathrm{\δ}}_{ij}^2}\]}{{\Σ}_{k=1}^u\exp \[-{\Σ}_{i=1}^r\frac{{\left(x_i-c_{ij}\right)}^2}{2{\mathrm{\δ}}_{ij}^2}\]};$

(4) in weighted mean layer, each neuronic output is that normalization layer Output rusults is multiplied by weighted deviation w ₂, this layer of neuron Output rusults is f _j=w ₂Ψ _j.

(5) each neuron in output layer represents the variable that the Output rusults by adding up in weighted mean layer obtains, and therefore, the Output rusults y of this layer is:

$y=\underset{j=1}{\overset{u}{\Σ}}{f}_j=\underset{j=1}{\overset{u}{\Σ}}{w}_2{\mathrm{\Ψ}}_j=\underset{j=1}{\overset{u}{\Σ}}{w}_z\frac{\exp \[-{\Σ}_{i=1}^r\frac{{\left(x_i-c_{ij}\right)}^2}{2{\mathrm{\δ}}_{ij}^2}\]}{{\Σ}_{k=1}^u\exp \[-{\Σ}_{i=1}^r\frac{{\left(x_i-c_{ij}\right)}^2}{2{\mathrm{\δ}}_{ij}^2}\]};$

The learning process of Self-organized Fuzzy Neural Network (SOFNN) mainly comprises parameter learning and Structure learning.

Parameter learning makes network Fast Convergent by online Recursive Least Squares.

Structure learning, by automatically increasing, revising or deleting the neuronic self-organization in ellipsoidal function layer, finds scale of neural network the most suitable.

Structure learning mainly comprises three committed steps:

1) neuron is increased;

2) neuron is pruned;

3) membership function in ellipsoidal function layer and fuzzy rule is merged.

Therefore, by the parameter learning in SOFNN and structure learning algorithm, optimum network structure can be found for each cluster and predict.

Embodiment

The present embodiment with somewhere 533 family resident for object carries out instance analysis, because this area is 11 at day of checking meter, we are using the electricity consumption data on January 10th, 11 days 1 April in 2014 as training set, and on January 11st, 2015 tests to the electricity consumption data on February 10 as test set.

Data preparation stage mainly comprises:

(1) data prediction: because the residential electricity consumption data of collected intelligent electric meter record are aggregate-value, therefore, calculates each user when the electricity consumption aggregate-value of daily power consumption needs by the electricity consumption aggregate-value on the same day being deducted proxima luce (prox. luc);

(2) missing values process: need after pre-service to detect the phenomenon that whether there is disappearance in data.By calculating before electricity consumption missing time section the difference of accumulative electricity one day after, and be averaged according to the number of days of disappearance, by the Data-parallel language of disappearance;

(3) outlier processing: the data sample of the metrics-thresholds scope exceeding setting is filtered, such as: the resident family etc. of street lamp, industrial user and long-term dereliction.

FCM cluster analysis

Traditional FCM algorithm, needs user to determine cluster number in advance.In cluster analysis, the determination of cluster number will produce a very large impact cluster result.In order to determine suitable cluster number objectively, the present embodiment, mainly through calculating error sum of squares (SSE) and MeanIndexAdequacy (MIA) value of each exploration classification number, finds out optimum cluster number parameter by comparative analysis.SSE and MIA computing formula is as follows:

$SSE=\underset{c=1}{\overset{c}{\Σ}}\underset{k=1}{\overset{n_c}{\Σ}}\left|\right|z_{ck}-v_c|{|}^2;$

$MIA=\sqrt{\frac{1}{C}\underset{c=1}{\overset{C}{\Σ}}\[\frac{1}{n_c}\underset{k=1}{\overset{n_c}{\Σ}}{(z_{ck}-v_c)}^2\]};$

Wherein, n _crepresent the number of the sample data in c class, z _ckrepresent the kth sample in c class.Choose optimum clustering number according to index changing tendency, the cluster result obtained, as shown in Figure 3, when class number is more than 6, along with the increase of class number, curve is more and more smooth, and the reduction trend of SSE and MIA desired value obviously weakens.

Meanwhile, in order to ensure that each cluster centre all has the sample of some, therefore, cluster number is set as 6, and cluster result is as shown in table 1.

Table 1 six cluster centre results

Cluster	Quantity	Average daily power consumption	Second ladder ratio	3rd ladder ratio	Sensitive
						1	75	0.0572	0.0277	0.0046	0.3535
2	113	0.1616	0.3794	0.0931	0.6148
						3	81	0.1965	0.8253	0.0712	0.4212
4	118	0.2944	0.5457	0.4166	0.5946
						5	89	0.3384	0.3976	0.5485	0.5988
6	57	0.4695	0.0645	0.9272	0.5224

Observe the numerical value of K-means cluster centre, in conjunction with the sample characteristics of each class, can following characteristics be summed up:

(1) the average power consumption of first kind user is minimum, and each class power consumption increases successively, and the average power consumption of the 6th class user is maximum.

(2) first kind user power utilization amount rests within the power consumption that the first step price specifies substantially, and the power consumption arriving second, third ladder is little; Equations of The Second Kind user power utilization amount is evenly distributed within the power consumption that first, second step price specifies; 3rd class user power utilization amount rests within the power consumption that the second step price specifies substantially, and the power consumption of the first, the 3rd ladder is little; The user of fourth, fifth class then mainly stops in second and third ladder.Wherein, the power consumption of the 4th class user more often drops on the second ladder, and the 5th class user more often drops on the 3rd ladder; 6th class user power utilization amount rests within the power consumption that the 3rd step price specifies substantially, and the power consumption of first, second ladder is few.

(3) can find out, for the first kind, the 3rd class, the 6th class user, step price can not reduce its power consumption very ideally, reason is, the power consumption of this kind of people every month is comparatively stable, almost be within a certain ladder power consumption constant, the fine setting of step price can not cause them to the remarkable change of domestic load.For Equations of The Second Kind, the 4th class, the 5th class user, step price may be obvious on its impact, and reason is that the power consumption of this three classes user is distributed in different step prices, generally there will be ladder span.When jumping into another ladder power consumption by a certain ladder power consumption, according to rationality economist hypothesis, the user that price sensitivity is high can correspondingly reduce its power consumption, avoids using more power consumption below higher electricity price.

Meanwhile, depict electricity consumption of resident curve family in each cluster respectively according to raw data set, as shown in Figure 4, therefrom can find that the power consumption of the user every day in classification 1 is all little, and very average, substantially concentrate on below 10kWh, insensitive to hot weather; User in classification 3 is insensitive to hot weather equally; And all the other classifications all summer power consumption have the increase of obvious amplitude, and the sensitive index of correspondence is higher, and the user in these classifications is more responsive for hot weather.

SOFNN load estimation

Residential electricity consumption data are added up according to cluster result, then erects corresponding SOFNN model respectively.Corrected by repeatedly simulated experiment and tracking error, estimate that preferably SOFNN preliminary experiment parameter is δ=0.01, σ ₀=0.1, krmse=0.01 and kd (i)=0.01 (i=1,2 ..., 8).

The accuracy of prediction is weighed by the mean absolute percentage error commonly used (MAPE) and is weighed with maximum absolute percentage error (ME), and computing formula is as follows:

$MAPE=\frac{{\Σ}_{n=1}^{\stackrel{\‾}{N}}\left|\frac{y_n-\widehat{y_n}}{y_n}\right|}{\stackrel{\‾}{N}}\×100\%;$

$ME=max\left|\frac{y_n-\widehat{y_n}}{y_n}\right|\×100\%;$

Wherein, y _nwith represent the actual value by this region electricity consumption total amount and predicted value, represent the number of days of prediction.

Corresponding to each cluster, the roll error of model is as shown in table 2.

The rolling forecast error of model corresponding to table 2 cluster

Error criterion

Model 1

Model 2

Model 3

Model 4

Model 5

Model 6

MAPE

22.87％

7.27％

5.14％

3.86％

2.79％

3.29％

ME

75.34％

17.43％

21.34％

9.80％

14.25％

8.90％

On the whole, except model 1, the rolling forecast error of each model all within the acceptable range.Reason is, there are some users in the cluster 1 corresponding to model 1, and their electricity consumption behavior randomness is strong especially, and its electricity consumption rule is difficult to screen.In classification forecast model, the electricity consumption behavior of this kind of user and other regular user power utilization behavior difference are comparatively large, often this kind of user are classified as a class.But due to features such as randomness are strong, the electricity consumption behavior of this kind of user often cannot accurately predicting.Fortunately, the power consumption that model 1 is predicted is less relative to the ratio of total consumption, also little on the precision of prediction impact of classification forecast model.

Each model prediction result is gathered, as shown in Figure 5, load estimation model result of finally being classified.Single load estimation model is respectively 3.34% and 2.78% with the MAPE value of classification load estimation model, and all within 4%, precision result is satisfactory.Compared to single load estimation model, the overall precision of prediction of classification load estimation model improves 0.56%.In addition, in the 14th day, (i.e. on January 25th, 2015), two model prediction accuracy dropped to less than 90% simultaneously, had consulted relevant historical data discovery, and its reason causes actual power consumption significantly to reduce suddenly because regionality has a power failure.For this accident, forecast model is difficult to carry out responding and revising to this in time, and this have impact on overall precision of prediction to a certain extent.

Above-described embodiment is only used to the present invention is described, and is not used as limitation of the invention.As long as according to technical spirit of the present invention, change above-described embodiment, modification etc. all will be dropped in the scope of claim of the present invention.

Claims (10)

1. a residential electricity consumption load predicting method in mid-term under step price mechanism, is characterized in that,

First, gather residential electricity consumption data, extract the attributive character of residential electricity consumption behavior under step price mechanism, identifying the different electricity consumption behavioural characteristic of resident under step price mechanism by cluster analysis, is same class of subscriber by the user Gui Ju with same or analogous electricity consumption behavioural characteristic;

Then, set up corresponding load estimation model respectively for each class of subscriber, and predict;

Finally, predicting the outcome of each class of subscriber is gathered.

2. residential electricity consumption load predicting method in mid-term under step price mechanism according to claim 1, it is characterized in that, first residential electricity consumption data are divided into several classes by different electricity consumption behavioural characteristics, obtain one group several there is the data set of different input and output, then corresponding load estimation model is set up to each data set.

3. residential electricity consumption load predicting method in mid-term under step price mechanism according to claim 1, is characterized in that, after gathering residential electricity consumption data, proceeds as follows:

1) data prediction: obtain each user and work as daily power consumption;

2) missing values process: if lack certain daily power consumption, then by calculating before electricity consumption missing time section the difference of accumulative electricity one day after, and be averaged according to the number of days of disappearance, as certain daily power consumption of disappearance;

3) outlier processing: the daily power consumption of working as of the metrics-thresholds scope exceeding setting is filtered.

4. residential electricity consumption load predicting method in mid-term under step price mechanism according to claim 1, it is characterized in that, when predicting, first the class of subscriber that user to be predicted returns will be identified, then corresponding load estimation model is selected to carry out load estimation, finally, gather single predicting the outcome, obtain final macro-forecast result.

5. residential electricity consumption load predicting method in mid-term under step price mechanism according to claim 1, is characterized in that, the attributive character extracting residential electricity consumption behavior comprises cluster Attributions selection, prediction input variable is extracted; Cluster attribute comprises the average daily power consumption of each user, the second ladder ratio, the 3rd ladder ratio and sensitive; Prediction input variable comprises the power consumption of seven days, the temperature on the same day in the past.

6. residential electricity consumption load predicting method in mid-term under step price mechanism according to claim 5, is characterized in that,

Average daily power consumption=total electricity consumption/total number of days of sampling;

Second ladder ratio=the arrive second ladder month number/total moon number;

3rd ladder ratio=arrival the 3rd ladder moon number/total moon number;

The average power consumption of sensitive=day with high temperature/average daily power consumption.

7. residential electricity consumption load predicting method in mid-term under step price mechanism according to claim 1, it is characterized in that, cluster analysis is realized by fuzzy c means means clustering algorithm, each attributive character of user is under the jurisdiction of one or more class of subscriber, represents that it belongs to the degree of different user classification with degree of membership.

8. residential electricity consumption load predicting method in mid-term under step price mechanism according to claim 1, it is characterized in that, load estimation model, based on Self-organized Fuzzy Neural Network model, comprises input layer, ellipsoidal function layer, normalization layer, weighted mean layer, output layer.

9. residential electricity consumption load predicting method in mid-term under step price mechanism according to claim 8, is characterized in that, the learning process of Self-organized Fuzzy Neural Network comprises parameter learning, Structure learning;

Parameter learning makes network Fast Convergent by online Recursive Least Squares;

Structure learning, by automatically increasing, revising or deleting the neuronic self-organization in ellipsoidal function layer, is searched and selects scale of neural network.

10. residential electricity consumption load predicting method in mid-term under step price mechanism according to claim 9, it is characterized in that, Structure learning comprises following operation:

1) neuron is increased;

2) neuron is pruned;

3) membership function in ellipsoidal function layer and fuzzy rule is merged.