CN103218675A - Short-term load prediction method based on clustering and sliding window - Google Patents

Abstract

The invention relates to a short-term load prediction method based on a clustering and sliding window. The method comprises the following steps of: preprocessing electric power load data; clustering historical data of a prediction user by utilizing a clustering algorithm, and adjusting clustering parameters; selecting k data from near to far of the prediction time in a category, containing most data, in clustering results to form a sliding window k; predicting the k selected data by utilizing a combination model based on the sliding window, and acquiring a primary prediction result; and correcting the primary prediction result of the combination model according to meteorological factors to obtain a final load prediction result. Compared with the prior art, the method has the advantages of high prediction precision, good adaptability and the like.

Description

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

Technical field

The present invention relates to the load forecast technical field, especially relate to a kind of short-term load forecasting method based on 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 of modern power systems and economical operation.Short-term electric load prediction is mainly used in forecast coming few hours, and whether it accurately was directly connected to the safe operation and the economic load dispatching of electric system until several days electric load in one day.High-precision short-term electric load prediction helps reasonably to arrange grid equipment scheduling and turnaround plan, improves the stability of Operation of Electric Systems, reduces the cost of electricity-generating of electrical network, helps improving the economic benefit and the social benefit of electric system.

The outstanding feature of short-term load forecasting is to be the similarity that the cycle presents variation with the day, obviously is subjected to the influence of weather conditions simultaneously.Therefore need fully research load variations rule, analysis load changes the relation of silver, particularly weather conditions, day type etc. and short-term load variations.At present,, mainly adopt methods such as time series, regretional analysis, trend extrapolation, gray model and neural network for short-term load forecasting, each tool relative merits of these algorithms, the scope of adaptation has nothing in common with each other.Because load prediction is subjected to the influence of a lot of uncertain factors, so far, there is not a kind of method to guarantee under any circumstance can both obtain satisfied predicting the outcome.

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

Summary of the invention

Purpose of the present invention is exactly to provide a kind of short-term load forecasting method based on cluster and moving window for the defective that overcomes above-mentioned prior art existence.

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

A kind of short-term load forecasting method based on cluster and moving window specifically may further comprise the steps:

Step 1: the electric load data are carried out pre-service, with of the requirement of adaptive clustering algorithm to data;

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

Step 3: in cluster result, comprise in the maximum class of data and choose k data from the near to the remote, constitute moving window k by the range prediction time;

Step 4: k data selecting are used predict, obtain and predict the outcome based on the built-up pattern of moving window;

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

Step 2 is specially:

Step 2.1: with the sky is the historical data that unit chooses certain predictive user, and classifies according to a Zhou Qitian;

Step 2.2: utilize clustering algorithm, data that will Monday to Sunday are cluster respectively, chooses respectively to comprise the maximum class of data.

Step 4 is specially:

Step 4.1: utilize different forecast models that k the data of choosing are predicted respectively;

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

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

The forecast model that is utilized in the 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, calculate each effective temperature constantly according to temperature, humidity and wind speed;

Step 5.2: in the moving window k that step 3 is chosen, set up the meteorologic factor correction model according to each load actual value, predicted value and effective temperature constantly;

Step 5.3: the meteorologic factor correction model of utilize determining, predicting the outcome of obtaining of combination forecasting revised, thereby obtained final predicting the outcome.

Compared with prior art, the present invention adopts clustering algorithm that the power load historical data is carried out cluster, excavated user's electricity consumption rule to the full extent, rejected the influence of small deviation load curve, comprise the maximum class of load curve after the cluster and carry out the forecast model training by choosing, can effectively improve the precision of short-term load forecasting.

Can guarantee to obtain the better prediction result in all cases without any a kind of Forecasting Methodology at present, predict based on the built-up pattern of moving window owing to adopt, fully utilize the information that various forecast model provides, can effectively improve the capability of fitting of forecast model.The weight of each forecast model is dynamically adjusted according to its predicated error, to improve precision of prediction and adaptivity.

In addition, the present invention has also adopted and can the different basic meteorological index effective temperatures of concentrated expression have been revised by predicting the outcome of load forecasting model of combination, with the influence of reflection weather conditions variation to short-term load forecasting, thereby further improves the load prediction precision.

Description of drawings

Fig. 1 is a process flow diagram of the present invention;

Fig. 2 is the process flow diagram of AP clustering algorithm.

Embodiment

The present invention is described in detail below in conjunction with the drawings and specific embodiments.

Embodiment

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

Step 1: the load data of gathering is carried out pre-service, form sample set, with of the requirement of adaptive clustering algorithm to data.

Because load data is the difference of adjacent two moment ammeter values, if load data is a negative value, ammeter reversing phenomenon promptly appears, this value is become null value.For missing values, utilize the trend penalty method, promptly numerical value power consumption was constantly arranged in the past according to the user, the historical electricity consumption trend of analysis user, and according to the same day ammeter value ratio fill null value.

After the load data pre-service, all data are carried out sample analysis, form sample set.Because all seven days load model differences, so separately set up sample, separately training, separately prediction.Simultaneously in order to guarantee that training sample is enough big, the collection of this example be integral point 24 hours every days load data constantly in two years.The data of need gathering except load data, also need to gather temperature (℃), relative humidity (%), wind speed weather datas such as (m/s).

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

Be that example describes with the Monday in the present embodiment, at the electricity consumption data of certain predictive user, with in history in 2 years the data of 104 Mondays carry out cluster, choose the maximum class of cluster result.Adopt the AP clustering algorithm to carry out cluster herein.

The AP clustering algorithm transmits Attraction Degree (responsiility) and degree of membership (availability) two class message.Wherein (i, k) expression reflects from the numerical value message that an i sends to candidate's cluster centre k whether the k point is suitable as the cluster centre that i is ordered to Attraction Degree message R.(i, k) expression reflects from the numerical value message that candidate's cluster centre k sends to i whether the i point selects k as its cluster centre to degree of membership message A.The AP algorithm is brought in constant renewal in the Attraction Degree and the degree of membership value of each point by iterative process, up to producing m high-quality cluster centre (exemplar), simultaneously remaining data point is assigned in the corresponding cluster.

Upgrade the Attraction Degree parameters R in iteration _fWith the degree of membership parameter A _fThe time, be to suppress the vibration influence, introduce ratio of damping λ, shown in formula (1) and (2), Attraction Degree parameters R wherein _tWith the degree of membership parameter A _tOnly be the nonce that calculates to formula (5) by formula (3), the final Attraction Degree parameters R that obtains of each iteration _fWith the degree of membership parameter A _fEqual Attraction Degree parameter nonce R _tWith degree of membership parameter nonce A _tWith the final Attraction Degree parameters R that obtains of last iteration _F-1With the degree of membership parameter A _F-1Weighted sum.

R _f＝(1-λ)*R _t+λ*R _f-1 (1)

A _f＝(1-λ)*A _t+λ*A _f-1 (2)

Wherein λ ∈ [0.5,1).

R _t(i, k) and A _t(i, computing formula k):

(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, AP clustering algorithm applicating flow chart as shown in Figure 2, step is as follows:

Step 2.1: for the sample set of Monday, calculate the Euclidean distance between sky and day load respectively in each sample set, negative Euclidean distance is the similarity between certain two days the most, forms 10816 * 10816 similarity matrix S.

In this example, for control final clusters number can be not excessive and class in data can be not very little, S (k in similarity matrix S, k) initial value can be made as the intermediate value of all values, finally by cluster result, constantly changes S (k, k) value is until satisfied cluster result occurring.Through test, (k k) gets 4 times of all values intermediate value in the matrix to S in this example.

Step 2.2: calculate Attraction Degree R and degree of membership A according to formula (1) to formula (5), sample set is carried out cluster.Therefrom choose the maximum class of clusters number respectively, the class that cluster is maximum has been rejected outlier, has reflected user's normal electricity consumption rule.

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

Step 4: 30 days the data use of selecting is predicted based on the built-up pattern of moving window, obtained preliminary predicting the outcome, specifically adopt following mode to carry out.

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

In the SVR of this example algorithm, at first utilize particle cluster algorithm to determine loss factor e and penalty factor c, utilize particle cluster algorithm than simple artificial definite parameter, improved accuracy.Predict according to 30 training samples by the SVR algorithm then.

The ARMA algorithm is regarded the forecasting object value as a random series that changes in time and constantly change, and this random series has embodied the continuity of predicted data.Obtain another predicted value by the ARMA algorithm predicts.

Step 4.2: utilize formula 6 to calculate SVR and the ARMA algorithm average relative error in moving window 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: calculate the weight of two kinds of prediction algorithm models respectively according to formula (7)-(8).

When t=1,

Along with the expansion of prediction, ω _iBy the average relative error of i kind model during moving window k

Draw.Computing formula is 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 ω _iDefinition as can be known, this built-up pattern is given big weight to the less model of error, and gives less weights to the bigger model of error, gives the maximum weights of predicated error least model, improve the contribution degree of this forecast model in finally predicting the outcome, to improve precision of prediction and confidence level.Along with new prediction arrival constantly, based on predetermined window the weights of each submodel are carried out update calculation, guaranteeing the dynamic self-adapting of load prediction, thereby improve the load prediction precision.

Step 4.4: according to two kinds of prediction algorithm predicted results and weight, according to formula (9) calculation combination predicted results.

If the individual predictor model of m (m 〉=2) is arranged, y in the combination load forecasting model _i(t) be i model (i=1...m) t load forecast value constantly, y (t) finally predicts the outcome for t moment 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 that choose, utilize meteorological correction model to revise, thereby draw final predicting the outcome, specifically adopt following mode to carry out to predicting the outcome.

Step 5.1: according to the temperature in 30 days that choose (℃), relative humidity (%), wind speed (m/s), utilize formula 10, calculate effective temperature respectively.Select effective temperature that short-term load forecasting is returned correction, be based on the strong correlation of load and effective temperature and fixed.

Effective temperature T _eComputing formula as follows:

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

T wherein _a, R _h, V be respectively temperature (℃), relative humidity (%), wind speed meteorologic parameter values such as (m/s).

Step 5.2: based on 30 days electric power load prediction values, actual value and effective temperatures choosing, (wherein T is effective temperature to use the least square regression method to obtain the regression parameter k of regression equation p '=kT+p+b and b, p is a predicted value, p ' is an actual value), obtain the short-term electric load prediction correction model shown in the formula (11) based on effective temperature:

p ₁＝kT+p+b (11)

Step 5.3:, utilize the meteorologic factor correction model to calculate final predicted value according to the predicted value of built-up pattern and the effective temperature of pre-observation.

And, can use the respective sample collection respectively from step 2 to step 5 for all the other six days in the week, thus draw the predicted value in a following week, thus complete load prediction finished.

Claims (5)

1. the short-term load forecasting method based on cluster and moving window is characterized in that, specifically may further comprise the steps:

Step 1: the electric load data are carried out pre-service, with of the requirement of adaptive clustering algorithm to data;

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

Step 3: in cluster result, comprise in the maximum class of data and choose k data from the near to the remote, constitute moving window k by the range prediction time;

Step 4: k data selecting are used predict, obtain and predict the outcome based on the built-up pattern of moving window;

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

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

Step 2.1: with the sky is the historical data that unit chooses certain predictive user, and classifies according to a Zhou Qitian;

Step 2.2: utilize clustering algorithm, data that will Monday to Sunday are cluster respectively, chooses respectively to comprise the maximum class of data.

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

Step 4.1: utilize different forecast models that k the data of choosing are predicted respectively;

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

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

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

5. a kind of short-term load forecasting method based on cluster and moving window according to claim 1 is characterized in that step 5 is specially:

Step 5.1: in the moving window k that step 3 is chosen, calculate each effective temperature constantly according to temperature, humidity and wind speed;

Step 5.2: in the moving window k that step 3 is chosen, set up the meteorologic factor correction model according to each load actual value, predicted value and effective temperature constantly;

Step 5.3: the meteorologic factor correction model of utilize determining, predicting the outcome of obtaining of combination forecasting revised, thereby obtained final predicting the outcome.

Images