CN107944594A

CN107944594A - One kind is based on SPSS and RKELM microgrid short-term load forecasting methods

Info

Publication number: CN107944594A
Application number: CN201710957155.8A
Authority: CN
Inventors: 杨荣照; 杨苹; 陈夏; 沈志钧; 余伟洲; 张云飞; 陈亦平; 张勇; 候君; 莫维科; 高琴
Original assignee: South China University of Technology SCUT
Current assignee: South China University of Technology SCUT
Priority date: 2017-09-30
Filing date: 2017-10-16
Publication date: 2018-04-20
Anticipated expiration: 2037-10-16
Also published as: CN107944594B

Abstract

The present invention proposes one kind based on SPSS and RKELM microgrid short-term load forecasting methods, including step：（1）On-line data acquisition simultaneously periodically updates historical data base（2）Historical data is pre-processed and extracts load sample feature；（3）Build offline load forecasting model；（4）The historical sample similar to be predicted forerunner's load is screened as on-line training sample using the method for Spearman rank correlation；（5）According to on-line training sample and offline load forecasting model, the predicted load of future time instance is calculated.This method simplifies kernel function extreme learning machine using quick（RKELM）, Chaos particle swarm optimization algorithm and Spearman rank correlation screening（RKELM）, establish comprising offline parameter optimizing and the prediction model in specific electric load；More newly arrived by cycle of model parameter and ensure the timeliness of algorithm, while reduce the complexity of on-line prediction calculating, reduction lasts data storage capacity, reduces and calculates cost, can accurately predict that microgrid is short-term and super short period load.

Description

One kind is based on SPSS and RKELM microgrid short-term load forecasting methods

Technical field

It is more particularly to a kind of to be born in short term based on SPSS and RKELM microgrids the invention belongs to microgrid load prediction technical field Lotus Forecasting Methodology

Technical background

Micro-capacitance sensor refers to by groups such as distributed generation resource, energy storage device, energy conversion device, load, monitoring and protective devices Into small-sized electric system.By the controllability of power supply and load inside micro-capacitance sensor, realize micro-capacitance sensor be incorporated into the power networks or Independent operating.Micro-capacitance sensor externally shows as an integral unit, while can smoothly be incorporated to major network operation again, meets user to electricity The requirement of energy quality and power supply safety.With the fast development of new energy new technology, distributed power generation progressively promotes and applies.Micro- electricity Net can promote distributed clean energy resource access, reduce environmental pollution and reduce power delivery loss, switched by orphan/grid-connect mode Improve the power supply reliability of user.Meanwhile the load prediction of microgrid is the important component of intelligent grid Energy Management System, Effective development of the load forecasting model to demand side management and intelligent power grid technology is most important.

Found by data analysis, relative to bulk power grid, the load randomness higher of micro-capacitance sensor, lasts data similarity not Greatly, predict that difficulty is larger.For ensure the efficient economical operation of micro-capacitance sensor, exactly load prediction be micro-capacitance sensor optimization operation and The important evidence of energy management decision-making.Mainly have to the method for the short-term load forecasting of electric system extrapolation, time series method, Artificial neural network method, support vector machines method etc..But microgrid load prediction of these methods for load randomness higher is not It is well suited for or to calculate cost higher.The present invention proposes that one kind based on SPSS and RKELM microgrid short-term load forecasting methods, passes through The methods of dimensionality reduction and Spearman rank correlation to kernel function screen (SPSS), while calculating cost is reduced, improves Certain precision, easy to being realized in common computer or built-in terminal device.

The content of the invention

In view of the problems of the existing technology the present invention, proposes one kind based on SPSS and RKELM microgrid short-term load forecastings Method, this method excavate the rule in data from demand history data, generate the extreme learning machine for simplifying kernel function (RKELM), the methods of and screening (SPSS) by Spearman rank correlation, while reducing calculating cost, has higher Prediction accuracy.

The solution of the present invention：One kind is included the following steps based on SPSS and RKELM microgrid short-term load forecasting methods：

Step A：On-line data acquisition periodicity simultaneously updates historical data base.

Step B：Historical data is pre-processed and extracts load sample feature.

Step C：Build offline load forecasting model：Using Chaos particle swarm optimization algorithm with simplifying kernel function extreme learning machine (RKELM) off-line training is carried out to demand history data, generates offline load forecasting model.

Step D：Using Spearman rank correlation (SPSS) method screening to be predicted forerunner's load is similar goes through History sample is as on-line training sample.

Step E：On-line prediction：According to on-line training sample and offline load forecasting model, the load of future time instance is calculated Predicted value.

Further, for the step A, on-line data acquisition is periodically and renewal lasts database, it is characterised in that bag Include：

Online acquisition data, and historical data base is updated daily, the data of 30 days before database preservation.

Further, for the step B, historical data is pre-processed and extracts load sample feature, its feature exists In, including：

Step B1：Historical data is normalized, zooms to [0,1] section, formula is as follows：

Wherein x_iThere is work value for i moment loads；x_minThere is work value for historical data minimum load in database；x_maxFor data Historical data peak load has work value in storehouse；For i moment load normalized values.

Step B2：Problem data P is labeled as to missing data or more than 4 continuous sampled values_k, repaiied according to the following formula Just：

Wherein P_k' it is revised data, P_k-96For the data (before i.e. one day) of 96 points before the k moment, P_k-192For the k moment The data of preceding 192 points, P_k-288For the data of 288 points before the k moment.The gathered data cycle is 15min.

Step B3：Extract load sample feature, including sampling time t_i, all information w_i, average load d a few days ago_ai, it is a few days ago stagnant Afterload d_li, week preceding hysteresis load w_li。

For average load d a few days ago_ai, its computational methods is shown below.

x_iThere are work value, x for i moment loads_i-jIt is that the loads of the i-th-j points has work value；da_iFor average load a few days ago, each Load point is corresponding with an average load a few days ago；Every 15 minutes points, average load is preceding 96 points of the load point a few days ago Average load；It is current load point that i, which is represented, and since the data of preceding 96 points do not have per day load, i is opened from 97 Begin, and j is exactly 1 to 96；G is the sampling sum of historical load sample；

For lagging load d a few days ago_liWith lagging load w before week_li, its computational methods is shown below.

dl_i=x_i-96；I=97 ... G

wl_i=x_i-672；I=673 ... G

For ease of statement, A is used_iAll load attributes of i-th of sample are referred to, its composition is shown below.

A_i={ t_i,w_i,da_i,dl_i,wl_i}

Further, for the step C, offline load forecasting model is built：Using Chaos particle swarm optimization algorithm with simplifying core Functional limit learning machine carries out off-line training to demand history data, generates offline load forecasting model.It is characterised in that it includes Following steps：

Step C1：Utilize sample M_TThe parameter to be optimized with RKELM algorithms (the extreme learning machine algorithm for simplifying kernel function)WithBuild the prediction model at T momentRecycle verification sample V_TLoad characteristic, be predicted and verify

In above formula, g trains algebraically for prediction model, corresponding with Chaos particle swarm optimization algorithm optimization algebraically；For the T moment The parameter of RKELM neural network models；For the gaussian kernel function parameter at T moment；For T moment RKELM neural network models Dimension；To verify the predicted load of sample；M_TFor the training sample at T moment, V_TFor the verification sample at T moment；A_iFor verification The feature of sample.

Step C2：Optimal model parameters, the target of Chaos particle swarm optimization algorithm (CPSO) are solved using chaotic particle algorithm group Function is minimum for prediction error, using mean absolute percentage error (Mean Absolute Percent Error) as pre- Evaluation function during model construction is surveyed, is shown below

S is predicted time overall length wherein in formula.Parameter optimization is carried out according to each group generation prediction error and its parameter

WhereinL∈[6,7,…,12]；Z₁,Z₂∈[-15,25]。

Step C3：Offline prediction model is built using optimized parameter C, γ and L.

Further, for the step D, using method screening and the to be predicted forerunner's load of Spearman rank correlation Similar historical sample is as on-line training sample.It is characterized in that, it is specially：

Take future position forerunner's load { x_i-1,x_i-2,x_i-3,…,x_i-16Be reference sample, arrange by size grade isCompare the historical load sample at preceding similar moment on the 8th and the similitude of reference sample, this Pierre Graceful coefficient of rank correlation is：

Wherein x '_i-mFor historical load sample.If the moment to be predicted is T °, 12 samples of ρ maximums are selected as online instruction Practice sample.

Further, for the step E, on-line prediction：According to on-line training sample and offline load forecasting model, calculate The predicted load of future time instance.It is characterized in that, it is specially：

12 samples filtered out by the use of Spearman rank correlation are as on-line training sample, with reference to offline load prediction Model, carries out on-line prediction.Meanwhile in order to ensure forecasting accuracy, the off-line optimization module interval phase is one week, using most The historical data operation of new 1 month, to obtain each moment optimized parameter C_T,γ_T,L_T(T=1,2 ..., 96), for next all On-line prediction module.

Compared with prior art, the present invention proposes that one kind optimizes protokaryon function model using RKELM models, and On the basis of to microgrid load data assay, using the method for Spearman correlated samples screening, precision is improved Computational efficiency is taken into account at the same time.

Brief description of the drawings

Fig. 1 is the optimized parameter compares figure that building micro-grid load predicts test in the 1st week in example.

Fig. 2 is the 1st week building micro-grid load prediction result figure in example.

Fig. 3 is the 1st week building micro-grid load Relative Error curve map.

Fig. 4 is one kind proposed by the present invention based on SPSS and RKELM microgrid short-term load forecasting method flow charts：

Fig. 5 realizes figure for microgrid load prediction principle.

Embodiment

It is described further below in conjunction with the accompanying drawings with implementation of the example to the present invention, but the implementation and protection of the present invention are unlimited In this.

One kind that Fig. 1 is the present invention is based on SPSS and RKELM microgrid short-term load forecasting methods.It is as shown in Figure 1, of the invention Method includes the following steps：

Step A：On-line data acquisition periodically simultaneously update historical data base, specifically include including：

By data acquisition device online acquisition data, and historical data base is updated daily, 30 days before database preservation Data.

Step B：Historical data is pre-processed and extracts load sample feature, it specifically comprises the following steps：

Inevitably there are wrong data or gaps and omissions in the data of collection, and therefore, it is essential that data are modified.

For average load d a few days ago_ai, its computational methods is shown below.

In formula：G is the sampling sum of historical load sample.

dl_i=x_i-96；I=97 ... G

wl_i=x_i-672；I=673 ... G

A_i={ t_i,w_i,da_i,dl_i,wl_i}

Step C：Build offline load forecasting model：Using Chaos particle swarm optimization algorithm with simplifying kernel function extreme learning machine (RKELM) off-line training is carried out to demand history data, generates offline load forecasting model.It is characterised in that it includes：

The neural network model of protokaryon functional limit learning machine (KELM) is shown below.

N is input layer dimension in formula.K(x_i,x_j) it is selected kernel function.Kernel function K (x_i,x_j), it usually takes Gaussian kernel letter Number, is shown below.

Ω_ELM(x_i,x_j)=K (x_i,x_j)=exp (- γ | | x_i-x_j||²)

It is a N × N-dimensional neutral net to understand KELM, and N is also training sample number at the same time.Obviously, under such structure, When N increases, network complexity increase, calculation amount can sharply increase, and which limits number of training, while limit kernel function The performance of extreme learning machine generalization ability, is unfavorable for the foundation of prediction model.

According to document (Deng Wanyu, Ong Yew Soon, Zheng Qinghua.A Fast Reduced Kernel Extreme Learning Machine [J] .NEURAL NETWORKS, 2016,76:29-38.), it is known that supporting vector can be with Machine is chosen, and it is RKELM that can simplify KELM accordingly.Its network is represented by：

In formula, β is the output weight of connection hidden layer and output layer, and y is RKELM training objective values, and L is RKELM nerves Network model dimension, N are input layer input data x_jDimension；

Being write as matrix form is then：

K_N×Lβ=Y

K in formula_N×L=k (X, X_L) it is simplified kernel matrix, β is the output weight vectors that length is L.

Minimize output weight beta:

In ELM algorithms, output weight beta usually takes its least square solution, and computational methods are shown below.

Based on above-mentioned formula, the neural network function of RKELM is represented by：

Thus, it is possible to it is neutral net that the neutral net of N × N-dimensional is reduced to N × L.Compared to KELM, RKELM is except C With outside γ, L is also required to Optimization Solution.

The selection mode of verification sample is that, if carrying out parameter optimization to the load forecasting model at T moment, verification sample is

V_T={ x_i|A_i}；t_i=T

In formula, V_TFor the verification sample of T moment load forecasting models；x_iFor predicted target values, namely actual value；A_iTo test Demonstrate,prove the feature of sample.

For training sample, according to forerunner's sample at T moment and preceding same period sample on the two, offline as the T moment is sought Training sample during optimal sorting, it is as follows

M_T={ x_i|A_i}；t_i=T-1, T-2,

T-94,…,T-98,T-192,…,T-196}

M in formula_TFor the training sample of T moment prediction models.

WhereinL∈[6,7,…,12]；Z₁,Z₂∈[-15,25]。

Chaos particle swarm optimization algorithm compares particle cluster algorithm, has the characteristics that ergodic, by will partly be absorbed in local extremum Particle go deactivation, that is, introduce chaos, make its jump out local extremum and improve obtain global optimum probability.Calculated with population The identical part of method is formula

WhereinRepresent that the g of i-th of particle subrogates to put, correspond to parameter Represent i-th particle The diverse vector of the g times, w are inertia coeffeicent, c₁,c₂Respectively degree of belief of the particle to itself and the degree of belief to colony, r₁, r₂For the random number between [0,1], pbestⁱFor the optimal location of i-th of particle, gbest is colony's optimal location.

When meeting following formula in the fitness function of particle continuous 5 generation

Judge that particle is absorbed in dead state, into Chaos Search pattern.Wherein per is threshold value.Randomly generate a 3-dimensional to Quantity space

Y₀={ Y_0,1,Y_0,2,Y_0,3}；Y_0,1,Y_0,2,Y_0,3∈[0,1]

With vectorial Y₀As iteration initial vector.According to Logistic equations.

Y_n+1=μ Y_n(1-Y_n)

Start chaos sequence iteration, obtain sequence of iterations Y₀,Y₁…Y_n-1, Logistic equations have the characteristics that ergodic, The multiple fields around locally optimal solution can be gone out with iteration.Then by way of carrier wave, according to

Y_i'=gbest+R (2Y_i-1)；I=0,1 ... n-1

It is amplified to centered on particle script position gbest, radius is that wherein R is Chaos Search radius on the region of R. Obtained Y_i' enter population searching process.When meeting required precision or reaching the iterations upper limit 200 times, ginseng is terminated Number searching process.

The off-line optimization module interval phase is one week, is run using the historical data of newest 1 month, to obtain each moment Optimized parameter C_T,γ_T,L_T(T=1,2 ..., 96), for next all on-line prediction modules.

Step C3：Utilize optimized parameter C_T,γ_T,L_T(T=1,2 ..., 96) the offline prediction model of structure.

Step D：Using Spearman rank correlation (SPSS) method screening to be predicted forerunner's load is similar goes through History sample is as on-line training sample.It is characterized in that, it is specially：

Step E：On-line prediction：According to on-line training sample and offline load forecasting model, the load of future time instance is calculated Predicted value.Specially：

Specific prediction example：

Herein using the historical load conduct of totally 70 days in resident building users mid-January~March of certain infrastructure management company collection Original sampling data；The sampling time interval of sample is 15min, and crude sampling attribute includes time and load power.

Before data import prediction model, all properties will be made with normalized, zoom ranges to [0,1], below with negative Exemplified by lotus power

The time-shared training of offline parameter optimizing, verification sample are used as by the use of the historical datas of first 40 days；The load of 30 days is gone through afterwards History data are the test sample of on-line prediction module, to test prediction error.

The accuracy estimating standard that load prediction uses is error：Missed herein using average relative error (MAPE), relatively Poor (APE)

Wherein S is forecast sample number；x_iFor test sample actual value；For test sample predicted value.

Offline parameter optimizing result

The parameter set for having RKELM each weeks of prediction, optimal value of the parameter when each parameter set includes 96 component are one day corresponding In from 0:00 starts every the at the time of point of 15min mono-.Exemplified by the 1st week of prediction test, each parameter value is as shown in Figure 1.

Load prediction results

It is similar to be predicted forerunner's load using the method screening of Spearman rank correlation after offline parameter optimizing Historical sample is obtained shown in load prediction results below figure 2 and Fig. 3 as on-line training sample：

From prediction result as can be seen that in most of the time section, load prediction results are more accurate, but also due to model is built Stand on the basis of similar day lags load data, when certain daily load prediction desired value is low with historical data similarity, prediction Value deviation is larger, as shown in the 6th day prediction result in Fig. 3.

Building load prediction error after the renewal of 1 model cycle of table

Table 1 is the 2nd~4 week, the prediction result error after the renewal of model optimized parameter cycle.It can be seen that this paper prediction models, Average relative error is predicted after being updated in the cycle no more than 12%, meets a month prediction error in industry and is not more than 15% It is required that.

The present invention has certain regularity but to change the characteristics of violent, diversity factor is big for building load, in kernel function and grain Improved on swarm optimization, to the reasonable dimensionality reduction of kernel function, Chaos Search is added in particle cluster algorithm so that model robustness Strengthened with generalization ability, and similar historical sample structure on-line load forcasting is rationally found based on Spearman rank correlation Model, improves flexibility and the predictive ability of prediction model.

Claims

1. one kind is based on SPSS and RKELM microgrid short-term load forecasting methods, it is characterised in that includes the following steps:

Step A：On-line data acquisition periodicity simultaneously updates historical data base；

Step B：Historical data is pre-processed and extracts load sample feature；

Step C：Build offline load forecasting model：Using Chaos particle swarm optimization algorithm with simplifying kernel function extreme learning machine (RKELM) off-line training is carried out to demand history data, generates offline load forecasting model；

Step D：The history sample similar to be predicted forerunner's load is screened using the method for Spearman rank correlation (SPSS) This is as on-line training sample.

Step E：On-line prediction：According to on-line training sample and offline load forecasting model, the load prediction of future time instance is calculated Value.

2. one kind according to claim 1 is based on SPSS and RKELM microgrid short-term load forecasting methods, it is characterised in that In the step A, including：

3. one kind according to claim 1 is based on SPSS and RKELM microgrid short-term load forecasting methods, it is characterised in that The step B includes：

<mrow> <mover> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>&OverBar;</mo> </mover> <mo>=</mo> <mfrac> <mrow> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>x</mi> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> </msub> </mrow> <mrow> <msub> <mi>x</mi> <mi>max</mi> </msub> <mo>-</mo> <msub> <mi>x</mi> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> </msub> </mrow> </mfrac> </mrow>

Wherein x_iThere is work value for i moment loads；x_minThere is work value for historical data minimum load in database；x_maxFor in database Historical data peak load has work value；For i moment load normalized values；

Step B2：Problem data P is labeled as to missing data or more than 4 continuous sampled values_k, according to the following formula amendment：

<mrow> <msub> <msup> <mi>P</mi> <mo>&prime;</mo> </msup> <mi>k</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mn>2</mn> </mfrac> <msub> <mi>P</mi> <mrow> <mi>k</mi> <mo>-</mo> <mn>96</mn> </mrow> </msub> <mo>+</mo> <mfrac> <mn>1</mn> <mn>3</mn> </mfrac> <msub> <mi>P</mi> <mrow> <mi>k</mi> <mo>-</mo> <mn>192</mn> </mrow> </msub> <mo>+</mo> <mfrac> <mn>1</mn> <mn>6</mn> </mfrac> <msub> <mi>P</mi> <mrow> <mi>k</mi> <mo>-</mo> <mn>288</mn> </mrow> </msub> </mrow>

Wherein P_k' it is revised data, P_k-96For the data of 96 points before the k moment, P_k-192For the number of 192 points before the k moment According to P_k-288For the data of 288 points before the k moment；The gathered data cycle is 15min；

Step B3：Extract load sample feature, including sampling time t_i, all information w_i, average load d a few days ago_ai, late negative a few days ago Lotus d_li, week preceding hysteresis load w_li；

For average load d a few days ago_ai, its computational methods is shown below：

<mrow> <msub> <mi>da</mi> <mi>i</mi> </msub> <mo>=</mo> <mfrac> <mn>1</mn> <mn>96</mn> </mfrac> <munderover> <mo>&Sigma;</mo> <mrow> <mi>j</mi> <mo>=</mo> <mn>1</mn> </mrow> <mn>96</mn> </munderover> <msub> <mi>x</mi> <mrow> <mi>i</mi> <mo>-</mo> <mi>j</mi> </mrow> </msub> <mo>;</mo> <mi>i</mi> <mo>=</mo> <mn>97</mn> <mo>...</mo> <mi>G</mi> <mo>,</mo> </mrow>

x_iThere are work value, x for i moment loads_i-jIt is that the loads of the i-th-j points has work value；da_iFor average load a few days ago, each load Point is corresponding with an average load a few days ago；Every 15 minutes points, average load is the flat of preceding 96 points of the load point a few days ago Equal load；It is current load point that i, which is represented, and since the data of preceding 96 points do not have per day load, i is since 97, and j It is exactly 1 to 96；G is the sampling sum of historical load sample；

For lagging load d a few days ago_liWith lagging load w before week_li, its computational methods is shown below：

dl_i=x_i-96；I=97 ... G

wl_i=x_i-672；I=673 ... G

For ease of statement, A is used_iAll load attributes of i-th of sample are referred to, its composition is shown below：

A_i={ t_i,w_i,da_i,d_li,w_li}。

4. one kind according to claim 1 is based on SPSS and RKELM microgrid short-term load forecasting methods, it is characterised in that In its described step C, including：

Step C1：Utilize sample M_TThe parameter to be optimized with RKELM algorithmsWithBuild the prediction model at T momentRecycle verification sample V_TLoad characteristic, be predicted and verify

<mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <msubsup> <mi>F</mi> <mi>T</mi> <mi>g</mi> </msubsup> <mo>=</mo> <mi>R</mi> <mi>K</mi> <mi>E</mi> <mi>L</mi> <mi>M</mi> <mrow> <mo>(</mo> <msubsup> <mi>C</mi> <mi>T</mi> <mi>g</mi> </msubsup> <mo>,</mo> <msubsup> <mi>&gamma;</mi> <mi>T</mi> <mi>g</mi> </msubsup> <mo>,</mo> <msubsup> <mi>L</mi> <mi>T</mi> <mi>g</mi> </msubsup> <mo>,</mo> <msub> <mi>M</mi> <mi>T</mi> </msub> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msubsup> <mover> <mi>x</mi> <mo>^</mo> </mover> <mi>i</mi> <mi>g</mi> </msubsup> <mo>=</mo> <msubsup> <mi>F</mi> <mi>T</mi> <mi>g</mi> </msubsup> <mrow> <mo>(</mo> <msub> <mi>A</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>,</mo> <msub> <mi>A</mi> <mi>i</mi> </msub> <mo>&Subset;</mo> <msub> <mi>V</mi> <mi>T</mi> </msub> </mrow> </mtd> </mtr> </mtable> </mfenced>

In above formula, g trains algebraically for prediction model, corresponding with Chaos particle swarm optimization algorithm optimization algebraically；For T moment RKELM The parameter of neural network model；For the gaussian kernel function parameter at T moment；For T moment RKELM neural network model dimensions；To verify the predicted load of sample；M_TFor the training sample at T moment, V_TFor the verification sample at T moment；A_iTo verify sample Feature.

Step C2：Optimal model parameters are solved using chaotic particle algorithm group, the object function of Chaos particle swarm optimization algorithm is prediction Error is minimum, is built using mean absolute percentage error (Mean Absolute Percent Error) as prediction model When evaluation function, be shown below

<mrow> <msup> <mi>MAPE</mi> <mi>g</mi> </msup> <mo>=</mo> <mfrac> <mrow> <mn>100</mn> <mi>%</mi> </mrow> <mi>S</mi> </mfrac> <munderover> <mo>&Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>S</mi> </munderover> <mrow> <mo>(</mo> <mo>|</mo> <mfrac> <mrow> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mover> <mi>x</mi> <mo>^</mo> </mover> <mi>i</mi> </msub> </mrow> <msub> <mi>x</mi> <mi>i</mi> </msub> </mfrac> <mo>|</mo> <mo>)</mo> </mrow> </mrow>

S is predicted time overall length wherein in formula；Parameter optimization is carried out according to each group generation prediction error and its parameter

<mrow> <mo>(</mo> <msubsup> <mi>C</mi> <mi>T</mi> <mrow> <mi>g</mi> <mo>+</mo> <mn>1</mn> </mrow> </msubsup> <mo>,</mo> <msubsup> <mi>&gamma;</mi> <mi>T</mi> <mrow> <mi>g</mi> <mo>+</mo> <mn>1</mn> </mrow> </msubsup> <mo>,</mo> <msubsup> <mi>L</mi> <mi>T</mi> <mrow> <mi>g</mi> <mo>+</mo> <mn>1</mn> </mrow> </msubsup> <mo>)</mo> <mo>=</mo> <mi>C</mi> <mi>P</mi> <mi>S</mi> <mi>O</mi> <mo>(</mo> <msup> <mi>MAPE</mi> <mi>g</mi> </msup> <mo>,</mo> <msubsup> <mi>C</mi> <mi>T</mi> <mi>g</mi> </msubsup> <mo>,</mo> <msubsup> <mi>&gamma;</mi> <mi>T</mi> <mi>g</mi> </msubsup> <mo>,</mo> <msubsup> <mi>L</mi> <mi>T</mi> <mi>g</mi> </msubsup> <mo>)</mo> </mrow>

WhereinL∈[6,7,...,12]；Z₁,Z₂∈[-15,25]；

5. one kind according to claim 1 is based on SPSS and RKELM microgrid short-term load forecasting methods, it is characterised in that In its described step D：

<mrow> <mi>&rho;</mi> <mo>=</mo> <mn>1</mn> <mo>-</mo> <mfrac> <mrow> <mn>6</mn> <munderover> <mo>&Sigma;</mo> <mrow> <mi>m</mi> <mo>=</mo> <mn>1</mn> </mrow> <mn>16</mn> </munderover> <msup> <mrow> <mo>(</mo> <msub> <mi>d</mi> <msub> <mi>x</mi> <mrow> <mi>i</mi> <mo>-</mo> <mi>m</mi> </mrow> </msub> </msub> <mo>-</mo> <msub> <mi>d</mi> <msubsup> <mi>x</mi> <mrow> <mi>i</mi> <mo>-</mo> <mi>m</mi> </mrow> <mo>&prime;</mo> </msubsup> </msub> <mo>)</mo> </mrow> <mn>2</mn> </msup> </mrow> <mrow> <msup> <mn>16</mn> <mn>3</mn> </msup> <mo>-</mo> <mn>16</mn> </mrow> </mfrac> </mrow>

Wherein x '_i-mFor historical load sample；If the moment to be predicted is T^o, 12 samples of ρ maximums are selected as on-line training sample This.

6. one kind according to claim 1 is based on SPSS and RKELM microgrid short-term load forecasting methods, it is characterised in that In its described step E：

By the use of 12 samples that Spearman rank correlation filters out as on-line training sample, with reference to offline load prediction mould Type, carries out on-line prediction；Meanwhile in order to ensure forecasting accuracy, the off-line optimization module interval phase is one week, and use is newest The historical data operation of 1 month, to obtain each moment optimized parameter C_T,γ_T,L_T(T=1,2 ..., 96), for next all On-line prediction module.