CN103218675B - A kind of based on the short-term load forecasting method of cluster and moving window - Google Patents

Abstract

The present invention relates to a kind of based on the short-term load forecasting method of cluster and moving window, comprise the following steps: electric power load data is carried out pre-treatment; Use cluster algorithm that the historical data of prediction user is carried out cluster, adjustment clustering parameter; Cluster result comprises in the maximum class of data by the range prediction time by closely to far choosing k certificate, form moving window k; The k selected is predicted based on the built-up pattern of moving window according to using, obtains tentative prediction result; According to meteorological factor, the tentative prediction result of built-up pattern is revised, obtain final load prediction results. Compared with prior art, the present invention has the advantages such as prediction precision height, adaptivity are good.

Description

A kind of based on the short-term load forecasting method of cluster and moving window

Technical field

The present invention relates to Electric Load Forecasting survey technology field, especially relate to a kind of based on the short-term load forecasting method of cluster and moving window.

Background technology

Short-term load forecasting is an important process of power department, is the important component part of energy management system, plays an important role in the safety and economical operation of modern power systems. Short-term electric load prediction is mainly used in forecast coming few hours, and within one day, until the electric power load of several days, whether it is accurately directly connected to safe operation of power system and economic scheduling. The short-term electric load prediction of high precision contributes to reasonably arranging grid equipment scheduling and turnaround plans, it is to increase the stability of Operation of Electric Systems, reduces the cost of electricity-generating of electrical network, is conducive to improving economic benefit and the social benefit of power system.

The outstanding feature of short-term load forecasting take day as the similarity cycle presenting change, simultaneously obviously by the impact of weather conditions. It is thus desirable to abundant research load variations rule, analysis load change silver, particularly weather conditions, day type etc. with the relation of short-term load variations. At present, for short-term load forecasting, mainly adopting the methods such as time series, regression analysis, trend extrapolation, grey model and neural network, each tool relative merits of these algorithms, the scope of adaptation is different. Owing to load prediction is subject to the impact of a lot of uncertain factor, so far, it does not have a kind of method ensures under any circumstance to obtain satisfied predicting the outcome.

In addition, in short-term load forecasting, historical data amount is excessive, and how choosing real effective historical data is also the problem needing primary study.

Summary of the invention

The object of the present invention is exactly the defect in order to overcome the existence of above-mentioned prior art and provides a kind of based on the short-term load forecasting method of cluster and moving window.

The object of the present invention can be achieved through the following technical solutions:

Based on a short-term load forecasting method for cluster and moving window, specifically comprise the following steps:

Step 1: electric power load data is carried out pre-treatment, with adaptive cluster algorithm to the requirement of data;

Step 2: use cluster algorithm that the historical data of prediction user is carried out cluster, adjustment clustering parameter, the maximum kind that cluster algorithm is obtained does not comprise the load data item that deviation is greater than threshold value, and the member of class is many as far as possible;

Step 3: comprise in the maximum class of data by the range prediction time by closely to far choosing k certificate in cluster result, form moving window k;

Step 4: the k selected being predicted based on the built-up pattern of moving window according to using, acquisition predicts the outcome;

Step 5: predicting the outcome of built-up pattern revised according to meteorological factor, obtains final load prediction results.

Step 2 is specially:

Step 2.1: the historical data choosing certain prediction user in units of sky, and classify according to a Zhou Qitian;

Step 2.2: utilize cluster algorithm, by the data on Monday to Sunday cluster respectively, chooses respectively and comprises the maximum class of data.

Step 4 is specially:

Step 4.1: utilize different predictive models respectively to the k chosen according to predicting;

Step 4.2: in the moving window k that step 3 is chosen, the average forecasting error in the window ranges of each predictive model in calculation procedure 4.1, and then calculate the weight of each predictive model in built-up pattern;

Step 4.3: according to each predictive model predict the outcome and its weight calculation in built-up pattern predicts the outcome.

The predictive model utilized in step 4.1 comprises arma modeling, supporting vector machine model and neural network model.

Step 5 is specially:

Step 5.1: in the moving window k that step 3 is chosen, calculates the true feeling temperature in each moment according to temperature, humidity and wind speed;

Step 5.2: in the moving window k that step 3 is chosen, be truly worth according to the load in each moment, predictor and true feeling temperature set up meteorological factor correction model;

Step 5.3: utilizing the meteorological factor correction model determined, predicting the outcome of combination forecasting being obtained is revised, thus obtains final predicting the outcome.

Compared with prior art, the present invention adopts cluster algorithm that power load historical data is carried out cluster, excavate the electricity consumption rule of user to the full extent, eliminate the impact of a small amount of deflection loads curve, carry out predictive model training by the class comprising load curve maximum after choosing cluster, can effectively improve the precision of short-term load forecasting.

Can ensure can obtain in all cases without any a kind of Forecasting Methodology at present and predict the outcome preferably, owing to adopting the built-up pattern based on moving window to predict, fully utilize the information that various predictive model provides, it is possible to effectively improve the capability of fitting of predictive model. The weight of each predictive model, according to its predicated error dynamic conditioning, predicts precision and adaptivity to improve.

In addition, the present invention additionally uses and can comprehensively reflect that predicting the outcome of combination load forecasting model is revised by different bases meteorological index true feeling temperature, to reflect that weather condition change is on the impact of short-term load forecasting, thus improves load prediction precision further.

Accompanying drawing explanation

Fig. 1 is the schema of the present invention;

Fig. 2 is the schema of AP cluster algorithm.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.

Embodiment

As shown in Figure 1, a kind of based on the short-term load forecasting method of cluster and moving window, specifically comprise the following steps:

Step 1: the load data gathered is carried out pre-treatment, forms sample set, with adaptive cluster algorithm to the requirement of data.

Owing to load data is the difference of adjacent two moment ammeter values, if load data is negative value, namely ammeter reversing phenomenon occurs, this value is turned into null value. For missing values, utilize trend compensation method, namely had the power consumption in numerical value moment according to user in the past, analyze user's history electricity consumption trend, and fill null value according to ammeter value ratio on the same day.

After load data pre-treatment, all data are carried out sample analysis, form sample set. Due to one week seven days load model differences, so separately setting up sample, separately training, separately prediction. Simultaneously in order to ensure that learning sample is enough big, the collection of this example be the load data in 24 hours integral point moment of every day in 2 years. Need the data gathered except load data, in addition it is also necessary to gather the weather datas such as temperature (DEG C), relative humidity (%), wind speed (m/s).

Step 2: use cluster algorithm that the historical data of prediction user is carried out cluster, adjustment clustering parameter, the maximum kind that cluster algorithm is obtained does not comprise the load data item that deviation is greater than threshold value, and the member of class is many as far as possible.

The present embodiment was described for Monday, for the electricity consumption data of certain prediction user, the data of 104 Mondays in history 2 years is carried out cluster, chooses the class that cluster result is maximum. AP cluster algorithm is adopted to carry out cluster herein.

AP cluster algorithm transmits Attraction Degree (responsiility) and degree of being subordinate to (availability) two class message. Wherein Attraction Degree message R (i, k) represents the numerical value message being sent to candidate cluster centre k from an i, and whether reflection k point is suitable as the cluster centre of i point. Degree of being subordinate to message A (i, k) represents the numerical value message being sent to i from candidate cluster centre k, and whether reflection i point selects k as its cluster centre. AP algorithm is constantly updated the Attraction Degree of each point by iterative process and is subordinate to angle value, until producing m high-quality cluster centre (exemplar), remaining data point is assigned in corresponding cluster simultaneously.

Attraction Degree parameter R is upgraded in iteration_fWith degree of being subordinate to parameter A_fTime, for suppressing vibration impact, introduce ratio of damping ��, as shown in formula (1) and (2), wherein Attraction Degree parameter R_tWith degree of being subordinate to parameter A_tIt is only the interim value calculated to formula (5) by formula (3), the Attraction Degree parameter R that each iteration finally obtains_fWith degree of being subordinate to parameter A_fEqual Attraction Degree parameter and it is worth R temporarily_tIt is worth A with degree of being subordinate to parameter temporarily_tWith the Attraction Degree parameter R that last iteration finally obtains_f-1With degree of being subordinate to parameter A_f-1Weighted sum.

R_f=(1-��) * R_t+��*R_f-1(1)

A_f=(1-��) * A_t+��*A_f-1(2)

Wherein �� �� [0.5,1).

R_t(i, k) and A_tThe calculation formula of (i, k):

R_t(i, k)=S (i, k)-max{A_f-1(i, j)+S (i, j) } (j �� 1,2 ..., N, but j �� k}) (3)

(4)

R_t(k, k)=P (k)-max{A_f-1(k, j)+S (k, j) } (j �� 1,2 ..., N, but j �� k}) (5)

In this example, as shown in Figure 2, step is as follows for AP cluster algorithm application schema:

Step 2.1: for the sample set of Monday, calculates the Europe formula distance between sky and sky load respectively in each sample set, and negative Europe formula distance is certain similarity between two days the most, the similarity matrix S of composition 10816 �� 10816.

In this example, in order to control, final clusters number can not data can not be very little in excessive and class, S (k in similarity matrix S, k) initial value can be set to the intermediate value of all values, eventually through cluster result, constantly change the value of S (k, k), until satisfied cluster result occurs. Through test, in this example, S (k, k) gets 4 times of all values intermediate value in matrix.

Step 2.2: calculate Attraction Degree R and degree of being subordinate to A according to formula (1) to formula (5), sample set is carried out cluster. Therefrom choosing the class that clusters number is maximum respectively, the class that cluster is maximum eliminates outlier, reflects the normal electricity consumption rule of user.

Step 3: moving window selects 30 days, namely certain the class data chosen in step 2 are from the close-by examples to those far off chosen the load data composing training sample of 30 days respectively, from sample, chooses temperature (DEG C), relative humidity (%), wind speed (m/s) accordingly simultaneously.

Step 4: use the built-up pattern based on moving window to predict the data of 30 days selected, obtain preliminary predicting the outcome, the following mode of concrete employing carries out.

Step 4.1: utilize different prediction algorithms that learning sample carries out training prediction respectively, utilize the ARMA algorithm in support vector regression (SVR) and time series in this example respectively, in actual applications, the not type of limit algorithm and quantity.

In the SVR algorithm of this example, first utilize particle cluster algorithm to determine loss factor e and punishment factor c, utilize the particle cluster algorithm artificially to determine parameter than simple, it is to increase accuracy. Then predicted according to 30 learning sample by SVR algorithm.

Forecasting object value is regarded as a stochastic sequence changed in time and constantly change by ARMA algorithm, and this stochastic sequence embodies the continuity of predicted data. Another predictor is obtained by ARMA algorithm predicts.

Step 4.2: utilize formula 6 to calculate SVR and ARMA algorithm in the average relative error of moving window in 30 days respectively.

$\left|\stackrel{\‾}{e_i}\right|=\frac{1}{k}\underset{j=1}{\overset{k}{\mathrm{\Σ}}}\left|e_i^j\right|=\frac{1}{k}\underset{j=1}{\overset{k}{\mathrm{\Σ}}}|p_i^j\left(\mathrm{predict}\right)-p_i^j\left(\mathrm{actual}\right)|---\left(6\right)$

Step 4.3: the weight calculating two kinds of prediction algorithm models according to formula (7)-(8) respectively.

As t=1,Along with the expansion of prediction, ��_iBy the average relative error of i-th kind of model during moving window kDraw. Calculation formula is as shown in (9)-(11).

$\left\{\begin{array}{cc}{\mathrm{\ω}}_1={\mathrm{\ω}}_2={\mathrm{\ω}}_3=...={\mathrm{\ω}}_m=\frac{1}{m}& t=1\\ {\mathrm{\ω}}_i=\frac{1/\left|\stackrel{\‾}{e_i}\right|}{\underset{i=1}{\overset{m}{\mathrm{\Σ}}}(1/|\stackrel{\‾}{e_i}\left|\right)}& t\≠1\end{array}---\left(7\right)\right.$

$\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i=1({\mathrm{\ω}}_1\≥0)---\left(8\right)$

By ��_iDefinition known, the model that error is less is given bigger weight by this built-up pattern, and the model that error is bigger is given less weights, gives the maximum weights of the minimum model of predicated error, improve the contribution degree of this predictive model in finally predicting the outcome, to improve prediction accuracy and confidence. Along with the arrival of new prediction time, carry out upgrading calculating to the weights of each submodel based on predetermined window, with the dynamic adaptivity of guaranteed load prediction, thus improve load prediction precision.

Step 4.4: predicting the outcome and weight according to two kinds of prediction algorithm models, according to predicting the outcome of formula (9) calculation combination model.

If combination load forecasting model having the individual prediction submodel of m (m >=2), y_iT () is the Electric Load Forecasting measured value of i-th model (i=1...m) t, y (t) finally predicts the outcome for t combination forecasting:

$y\left(t\right)=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i{y}_i\left(t\right)---\left(9\right)$

Step 5: according to the weather data in 30 days chosen, utilizes meteorological correction model to revise predicting the outcome, thus draws final predicting the outcome, and the following mode of concrete employing carries out.

Step 5.1: according to the temperature (DEG C) in 30 days chosen, relative humidity (%), wind speed (m/s), utilize formula 10, calculates true feeling temperature respectively. Select true feeling temperature to carry out short-term load forecasting returning and revise, it is the strong correlation based on load and true feeling temperature and fixed.

True feeling temperature T_eCalculation formula as follows:

$T_e=37-\frac{37-T_a}{0.68-0.14R_h+\frac{1}{1.76+1.4V^{0.75}}}-0.29T_a(1-R_h)---\left(10\right)$

Wherein T_a��R_h, V be respectively the meteorological parameter value such as temperature (DEG C), relative humidity (%), wind speed (m/s).

Step 5.2: based on 30 days electric power load predictors, actual value and the true feeling temperature chosen, (wherein T is true feeling temperature to the regression parameter k and b of use least-squares regression approach acquisition regression equation p '=kT+p+b, p is predictor, p ' is true value), obtain the short-term electric load prediction correction model based on true feeling temperature shown in formula (11):

p₁=kT+p+b (11)

Step 5.3: according to the true feeling temperature of the predictor of built-up pattern and prediction sky, utilize meteorological factor correction model to calculate final predictor.

And for all the other in a week six days, it is possible to use corresponding sample set from step 2 to step 5 respectively, thus draw the predictor of following a week, thus complete the prediction of complete load.

Claims (4)

1. one kind based on the short-term load forecasting method of cluster and moving window, it is characterised in that, specifically comprise the following steps:

Step 1: electric power load data is carried out pre-treatment, with adaptive cluster algorithm to the requirement of data;

Step 2: use cluster algorithm that the historical data of prediction user is carried out cluster, adjustment clustering parameter, the maximum kind that cluster algorithm is obtained does not comprise the load data item that deviation is greater than threshold value, and the member of class is many as far as possible;

Step 3: comprise in the maximum class of data by the range prediction time by closely to far choosing k certificate in cluster result, form moving window k;

Step 4: the k selected being predicted based on the built-up pattern of moving window according to using, acquisition predicts the outcome;

Step 5: predicting the outcome of built-up pattern revised according to meteorological factor, obtains final load prediction results,

Step 5 is specially:

Step 5.1: in the moving window k that step 3 is chosen, calculates the true feeling temperature in each moment according to temperature, humidity and wind speed:

$T_e=37-\frac{37-T_a}{0.68-0.14R_h+\frac{1}{1.76+1.4V^{0.75}}}-0.29T_a(1-R_h)$

Wherein: T_eFor true feeling temperature, T_a��R_h, V be respectively temperature, relative humidity, wind speed;

Step 5.2: in the moving window k that step 3 is chosen, be truly worth according to the load in each moment, predictor and true feeling temperature set up meteorological factor correction model:

p_t=kT_e+p+b

Wherein: p_tFor the predictor after correction, p is the predictor before revising, and k, b are regression parameter;

Step 5.3: utilizing the meteorological factor correction model determined, predicting the outcome of combination forecasting being obtained is revised, thus obtains final predicting the outcome.

2. according to claim 1 a kind of based on the short-term load forecasting method of cluster and moving window, it is characterised in that, step 2 is specially:

Step 2.1: the historical data choosing certain prediction user in units of sky, and classify according to a Zhou Qitian;

Step 2.2: utilize cluster algorithm, by the data on Monday to Sunday cluster respectively, chooses respectively and comprises the maximum class of data.

3. according to claim 1 a kind of based on the short-term load forecasting method of cluster and moving window, it is characterised in that, step 4 is specially:

Step 4.1: utilize different predictive models respectively to the k chosen according to predicting;

Step 4.2: in the moving window k that step 3 is chosen, the average forecasting error in the window ranges of each predictive model in calculation procedure 4.1, and then calculate the weight of each predictive model in built-up pattern;

Step 4.3: according to each predictive model predict the outcome and its weight calculation in built-up pattern predicts the outcome.

4. according to claim 3 a kind of based on the short-term load forecasting method of cluster and moving window, it is characterised in that, the predictive model utilized in step 4.1 comprises arma modeling, supporting vector machine model and neural network model.