CN103699947A - Meta learning-based combined prediction method for time-varying nonlinear load of electrical power system - Google Patents

Abstract

The invention discloses a meta learning-based combined prediction method for a time-varying nonlinear load of an electrical power system. The method has the beneficial effects that (1) meta learning is a final result obtained by learning for a plurality of times on the basis of the learning result, and an output result of the previous layer of model and a characteristic attribute of a prediction sequence are utilized as input information of the next layer of learning, so that the previous learning can be fully applied to the later conclusion process, so that system deviation in the used learning algorithm can be found out and corrected and the learning accuracy is improved; (2) the optimal weight is obtained by setting the mean square error to the minimum and adjusting a gating network parameter through a decision condition in meta learning. Thus, important reference basis is provided for optimization and determination of the weight of the load combined prediction model.

Description

During a kind of electric system based on unit's study, become nonlinear-load combination forecasting method

Technical field

The present invention relates to a kind ofly for Load Prediction In Power Systems method, belong to Power System and its Automation load prediction technical field.

Background technology

Electric power is the important foundation of national economy, is emphasis and look-ahead industry in national economic development strategy, and load prediction is as one of key job of electrical production department, for the normal operation of electric system provides condition.It reduces spinning reserve capacity for arranging economically Unit Commitment, and the plan of Optimum unit maintenance reduces cost of electricity-generating, increases economic efficiency, and has vital effect.And improve precision of prediction, be the precondition of giving full play to Load Prediction In Power Systems effect, there is important practical significance.

At present, conventional load forecasting method has: correlation predictive method, linear regression analysis predicted method, grey prediction, wavelet decomposition and the Reconstruction Method of the trend extrapolation predicted method based on similar day, time series forecasting, consideration meteorologic factor, support vector machine Regression Forecast, chaos forecast method, neural network prediction method etc.What above-mentioned Forecasting Methodology all adopted is Individual forecast model, causes some time period prediction deviation larger.

Because electric load has non-linear, time variation and probabilistic feature, the mathematical model applicability that these single Forecasting Methodologies adopt is limited, can not be in time, exactly the parameter of forecast model is carried out to Estimation and rectification, more be not easy to describe the unexpected variation of load, therefore there is certain limitation, be difficult to obtain higher precision of prediction.The present invention will predict that load sequence signature attribute is as weights basis, according to the minimum network parameter of adjusting of square error, try to achieve optimum weights, set up and have more high-precision load prediction built-up pattern.

Summary of the invention

Technical matters to be solved by this invention is for improving Load Prediction In Power Systems precision, and a kind of power system load combination forecasting method based on unit's study is provided.

The technical scheme that technical solution problem of the present invention adopts is as follows: a kind of power system load combination forecasting method based on unit's study, comprises the following steps:

(1): adopt the load in a certain moment in the available load forecasting method prediction of N kind section preset time, the load prediction results that obtains various Forecasting Methodologies is

form the base fallout predictor of unit's study;

(2): at first fallout predictor training stage characteristic attribute, can directly obtain, in first predictor predicts stage, adopt day typical curve to replace prediction curve at prediction characteristic attribute constantly at correspondence characteristic attribute constantly, build unit's study prediction characteristic attribute constantly.

(3): using base fallout predictor as unit's prediction input, using data characteristics attribute as gating network, input obtains the weight of each base fallout predictor, becomes Combination Forecasting model while building unit's study.

(4): according to the whole square error energy function E of sample _fi guarantee to predict the outcome and desired output matching optimum, thereby determine optimum network parameter.

Described available predictions model includes but not limited to following four kinds:

1) simple index number curve model: y ₀=c ₁+ x ⁿ

2) dispensable mould polynomial curve type: y ₁=a ₁+ b ₁x

3) quadratic form polynomial curve type: y ₂=a ₂+ b ₂x+c ₂x ²

4) cubic form polynomial curve type: y ₃=a ₃+ b ₃x+c ₃x ²+ d ₃x ³

A in formula _i, b _i, c _i, d ₃, n (i=1,2,3) is solve for parameter, by least square method, estimates to determine.

The corresponding characteristic attribute constantly of day typical curve adopting comprises following four characteristic attributes:

(1) the prediction moment and the front and back first order difference mean square deviation a that T is ordered altogether thereof _(p)(1);

(2) the prediction moment and the front and back second order difference mean square deviation a that T is ordered altogether thereof _(p)(2);

(3) the prediction moment and the front and back third order difference mean square deviation a that T is ordered altogether thereof _(p)(3);

(4) the prediction moment and the front and back chain rate coefficient mean square deviation a that T is ordered altogether thereof _(p)(4).

The concrete steps of step (3) are:

A, original input data X _pbe input to n ground level fallout predictor, obtain the ground level f that predicts the outcome ₁(X _p), f ₂(X _p) ..., f _n(X _p).By original input data proper vector a _pas gating network input, obtain the weight c of each ground level fallout predictor ₁(X _p), c ₂(X _p) ..., c _n(X _p).

B, in gating network, use softmax activation function, the weight c of i ground level fallout predictor _i(X _p) be:

$c_i\left(X_p\right)=\frac{e^{Z_i\left(a_p\right)}}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{e}^{\mathrm{Zi}\left(a_p\right)}}---\left(1\right)$

Wherein: $Z_i\left(a_p\right)=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_{\mathrm{ki}}\·a_p\left(k\right),i=\mathrm{1,2},\mathrm{\Λ},n.$

C, combined prediction finally predict the outcome into:

$F\left(X_p\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{f}_i\left(X_p\right)\·c_i\left(X_p\right)---\left(2\right)$

The concrete steps of step (4) are:

A, combined prediction result F (X _p) and corresponding desired output y _ptwo sequence matchings are optimum, can be by guaranteeing the whole square error energy function E of training sample _fminimum obtains:

$E_F=\frac{1}{P}\underset{p=1}{\overset{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HSA0000100402810000021.png"\; state="\{\{state\}\}">P</figure-callout>}}{\mathrm{\&Sigma;}}}\frac{1}{2}{(y_p-F\left(X_p\right))}^2---\left(3\right)$

Wherein: the number that P is training sample.Solve on this basis minimum E _fdetermine network parameter.

B, calculate instantaneous gradient vector

Wherein: ω _kifor network parameter, y _pfor desired output, a _p(k) be characteristic attribute.

C, determine network parameter

${\mathrm{\&omega;}}_{\mathrm{ki}}^{\mathrm{new}}={\mathrm{\&omega;}}_{\mathrm{ki}}^{\mathrm{old}}-\frac{\&PartialD;E_F}{\&PartialD;{\mathrm{\&omega;}}_{\mathrm{KI}}}---\left(5\right)$

Wherein:

for the network parameter of last iteration,

for the network parameter upgrading.

Beneficial effect of the present invention is as follows: the study of (1) unit is on the basis of learning outcome, repeatedly to learn to obtain net result.The input message that it utilizes the Output rusults of last layer model and the characteristic attribute of forecasting sequence to learn as lower one deck, make previous study can fully be applied to generalization procedure below, thereby find and correct the system deviation in used learning algorithm, improve study precision.(2) unit's study is minimum by setting square error, adjusts gating network parameter, thereby obtain optimum weights by decision condition, determines that the weight of Load forecasting provides reference frame for optimizing.

Accompanying drawing explanation

The power system load combination forecasting method flow process of Fig. 1 based on unit's study.

The combined prediction device structure of Fig. 2 based on unit's study.

The structure of Fig. 3 gating network.

The comparison that predicts the outcome of the combination forecasting of Fig. 4 based on unit study and neural network ensemble forecast model.

Fig. 5 is prediction load curve and realized load curve comparison diagram.

Fig. 6 is combined prediction load curve and realized load curve comparison diagram.

Specific implementation method

Below in conjunction with drawings and Examples, the present invention is described in further detail.But the invention is not restricted to given example.

According to Fig. 1 predict flow process adopt respectively curve model, conic model, cubic curve model, simple index number curve model to certain regional power grid on March 1st, 2004 to March 10, the load of totally 2880 is predicted, predicts the outcome as shown in Figure 5.Using predicting the outcome as the input of unit's prediction base fallout predictor, extract forecast period characteristic attribute, input gate network calculations combining weights obtains combined prediction result, and the combined prediction device structure of unit's study is as Fig. 2.Minimum according to whole square error energy function, gating network gain of parameter combined prediction optimal result is adjusted in circulation, and the adjustment of network parameter as shown in Figure 3.

Curve model, conic model, cubic curve model, a simple index number curve model are carried out to averagesofforecasts prediction, neural network ensemble prediction and the combined prediction result based on unit's study as shown in Figure 6.

The training sample adopting in curve model, conic model, cubic curve model, simple index number curve model, averagesofforecasts forecast model, neural network ensemble forecast model and the combination forecasting training process based on unit's study is 8928 training samples on July 1st, 2004 to July 31, and prediction average error result is as table 1.

Table 1

Table 1 is to adopt different Individual forecast methods and several common combinations Forecasting Methodology prediction average error, in reason table, result is visible, and predicting the outcome of the combination forecasting based on unit's study is better than predicting the outcome of various Individual forecast models and predicting the outcome of other common combinations forecast models.

Get predicting the outcome of on March 6th, 2007, as Fig. 4.As seen from the figure, the predicting the outcome of combination forecasting based on unit's study is better than predicting the outcome of neural network ensemble forecast model.

Table 2

Table 2 is three prediction moment in Fig. 4, while adopting the combination forecasting of learning based on unit to predict, and the weight of different base forecast models.12:30 is that the crest moment, 23:00 are the trough moment, and 17:55 is a moment between crest and trough.In table, 12:30 is loaded while predicting constantly, and curve model weight and conic model weight are all higher, though this is because be flex point in this moment, but variation is slower on part, therefore with a curve model, carries out prediction effect simultaneously and will get well; 23:00 is loaded while predicting constantly, and conic model weight is higher, be because be flex point constantly at this, and localized variation is than very fast, therefore by conic model, carries out prediction effect and will get well; 17:55 is loaded while predicting constantly, and one time curve model weight is higher, and this is because this is non-flex point constantly, therefore with a curve model, carries out prediction effect and will get well.Therefore the combination forecasting based on unit's study has guaranteed the time variation of combined prediction.

Each weight is all nonnegative number as seen from Table 2, and therefore the combination forecasting based on unit's study has guaranteed the nonnegativity of combined prediction.

Therefore by table 2, it can also be seen that simple index number curve model predicts the outcome bad, when each is predicted constantly, shared weight is all little.Thus, the visible combination forecasting of learning based on unit that adopts can effectively be rejected bad model, the choose reasonable of implementation model.

Claims (4)

1. become a nonlinear-load combination forecasting method during electric system based on unit study, it is characterized in that: comprise the following steps:

(1): chosen dispensable mould polynomial curve type, quadratic form polynomial curve type, cubic form polynomial curve type or simple index number curve model as base fallout predictor;

(2): in first predictor predicts stage, adopt day typical curve to replace prediction curve at prediction characteristic attribute constantly at correspondence characteristic attribute constantly;

(3): proper vector is inputted to gating network and obtain weight, while setting up unit's study, become Combination Forecasting model;

(4): according to the whole square error energy function E of sample _fminimum makes to predict the outcome optimum with desired output matching, determines suitable network parameter.

2. become nonlinear-load combination forecasting method during a kind of electric system based on unit study according to claim 1, it is characterized in that: the concrete steps of step (2) are:

After having chosen many similar day curves of prediction day, can form a day typical curve X _s, due to day typical curve X _sin general morphology, be and prediction day curve X _psimilar, therefore can adopt X _scharacteristic attribute be similar to and characterize prediction daily load curve X _pcharacteristic attribute;

For known sequence X _swhether there is polynomial curve trend and possess several rank polynomial curve trend, can be by sequence X _sdifferential Characteristics characterize, the first order difference of dispensable mould polynomial curve type is identical, the second order difference of conic model is constant, the third order difference of cubic curve model is constant; For given sequence X _sby the chain rate characteristic present index curve trend of sequence, the developmentspeed with linkrelative method of simple index number curve model is constant;

Select following four characteristic attributes:

(1) the prediction moment and the front and back first order difference mean square deviation a that T is ordered altogether thereof _(p)(1);

(2) the prediction moment and the front and back second order difference mean square deviation a that T is ordered altogether thereof _(p)(2);

(3) the prediction moment and the front and back third order difference mean square deviation a that T is ordered altogether thereof _(p)(3);

(4) the prediction moment and the front and back chain rate coefficient mean square deviation a that T is ordered altogether thereof _(p)(4).

3. become nonlinear-load combination forecasting method during a kind of electric system based on unit study according to claim 1, it is characterized in that: the concrete steps of step (3) are:

A, original input data X _pbe input to n ground level fallout predictor, obtain the ground level f that predicts the outcome ₁(X _p), f ₂(X _p) ..., f _n(X _p); By original input data proper vector a _pas gating network input, obtain the weight c of each ground level fallout predictor ₁(X _p), c ₂(X _p) ..., c _n(X _p).Combined prediction device finally predict the outcome into:

$F\left(X_p\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{f}_i\left(X_p\right)\&CenterDot;c_i\left(X_p\right)---\left(1\right)$

B, in gating network, use softmax activation function, the weight c of i ground level fallout predictor _i(X _p) be:

$c_i\left(X_p\right)=\frac{e^{Z_i\left(a_p\right)}}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{e}^{\mathrm{Zi}\left(a_p\right)}}---\left(2\right)$

Wherein: $Z_i\left(a_p\right)=\underset{k=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{\mathrm{ki}}\&CenterDot;a_p\left(k\right),i=\mathrm{1,2},\mathrm{\&Lambda;},n.$

4. become nonlinear-load combination forecasting method during a kind of electric system based on unit study according to claim 1, it is characterized in that: the concrete steps of step (4) are:

A, combined prediction result F (X _p) and corresponding desired output y _ptwo sequence matchings are optimum, can be by guaranteeing the whole square error energy function E of training sample _fminimum obtains:

$E_F=\frac{1}{P}\underset{p=1}{\overset{P}{\mathrm{\&Sigma;}}}\frac{1}{2}{(y_p-F\left(X_p\right))}^2---\left(3\right)$

Wherein: the number that P is training sample;

B, calculate instantaneous gradient vector

Wherein: ω _kifor network parameter, y _pfor desired output, a _p(k) be characteristic attribute;

C, modification network parameter

${\mathrm{\&omega;}}_{\mathrm{ki}}^{\mathrm{new}}={\mathrm{\&omega;}}_{\mathrm{ki}}^{\mathrm{old}}-\frac{\&PartialD;E_F}{\&PartialD;{\mathrm{\&omega;}}_{\mathrm{KI}}}---\left(5\right)$

Wherein:

for the network parameter of last iteration,

for the network parameter upgrading.

Images