CN106599417A - Method for identifying urban power grid feeder load based on artificial neural network - Google Patents

Abstract

The invention provides a method for identifying urban power grid feeder load based on an artificial neural network. The method comprises the following steps: (1) searching equipment composition, historical power consumption information and meteorological information of typical construction load of a corresponding climatic region; (2) establishing power consumption load models of all kinds of typical constructions; (3) generating load curves of corresponding time periods of all kinds of typical construction loads based on date and meteorological information of a to-be-identified feeder load curve; and (4) identifying the load composition of the to-be-identified feeder load curve by using a BP artificial neural network algorithm. According to the method provided by the invention, different construction load composition of large feeder nodes is effectively identified, which is conducive to predicting the dynamic load properties of the feeder nodes and evaluating the demand response potentials, and foundation is laid for regulating a reasonable demand response policy and mechanism.

Description

A kind of urban distribution network feeder load based on artificial neural network constitutes discrimination method

Technical field

The present invention relates to a kind of electrical network feeder load constitutes discrimination method, and in particular to a kind of based on artificial neural network Urban distribution network feeder load constitutes discrimination method.

Background technology

The composition of the large-scale feed connection node load of power system is related to terminal use's electricity consumption behavior characteristicss, cycle and custom, Always as season, date and external environment condition persistently change the dynamic electrical characteristics so as to affect whole system load and to electricity Net forms disturbance.The load of peak period is transferred to by the balance that low-valley interval is beneficial to hair electricity by dsm, But the planning of management policies and formulate such as dynamic electricity price Mechanism Design it will be clear that the composition of different supply node loads, so as to have Effect assessment feed connection node load flexibility and schedulable potentiality.

Chinese scholars have carried out numerous studies in terms of load structure identification, mainly include intrusive mood load monitoring (intrusive load monitoring, ILM) and non-intrusion type load monitoring (nonintrusive load monitoring,Non-ILM).ILM depends on user to equipment operation, electric quantity consumption, electric cost expenditure and intelligent electric meter The record data of the daily electricity consumption behavior such as situation, needs to install real-time watch device in user side.And it is all of different user It is very expensive that electrical equipment extensively installs test equipment, will also result in the dimension calamity of data storage and analysis.Non-ILM according to Rely the high-resolution in load signal, such as dynamic load response, instantaneous power or active change after current waveform, disturbance, extensively The non-intrusion type load monitoring being applied to when resident apartment or less electric power.However, for large-scale user or power system Load feeder node (electric power is often as high as MW or GW), it is this by detecting that the pattern of increment change is difficult to isolate generation The load type and equipment of change.

The content of the invention

In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a kind of urban distribution network based on artificial neural network and presents Specific electric load constitutes discrimination method, and the present invention effectively recognizes large-scale feed connection node difference building load and constitutes, contributes to predicting feeder line Node load dynamic characteristic, evaluation requirement response potentiality, are that the rational demand response policy of formulation and mechanism lay the foundation.

In order to realize foregoing invention purpose, the present invention is adopted the following technical scheme that：

A kind of urban distribution network feeder load based on artificial neural network constitutes discrimination method, and methods described includes：

(1) equipment composition, history power information and the weather information of correspondence climatic province typical building load are collected；

(2) the power load model of all kinds of typical buildings is set up；

(3) based on feeder load curve to be identified corresponding date and weather information, all kinds of typical building loads pair are generated Answer the load curve of period；

(4) identification of load structure is carried out to the feeder load to be identified using BP artificial neural network algorithms.

Preferably, in the step (3), the feeder load curve to be identified be by feeder line to be assessed week load curve and The load curve that load can be recognized subtracts each other acquirement.

Preferably, in the step (4), comprise the steps：

Step 4-1, netinit, to each connection weight initial value, computational accuracy and maximum study number of times are assigned, and arrange k= 0；

Step 4-2, all load curves of n quasi-representatives building for generating and corresponding weather information, using Monte Carlo sampling structure Into feed connection node week load curve, and form three layers of BP network structures, including input layer, hidden layer and output layer；

Step 4-3, using network desired output and reality output, the local derviation of each neuron of calculation error function pair output layer Number, the connection weight and the output of hidden layer using hidden layer to output layer, each neuron of calculation error function pair hidden layer Partial derivative；

Step 4-4, connection weight is corrected with the output of each neuron of output layer and each neuron of hidden layer, using implicit The input of each neuron of layer and each neuron of input layer is correcting connection weight；

Step 4-5, calculating global error；

If step 4-6, study number of times are more than or equal to maximum times m of setting, terminate calculating；If study number of times is less than The maximum times of setting, into step 4-7；

If step 4-7, network error meet required, terminate calculating；If be unsatisfactory for, k=k+1 goes to step 4-2.

Preferably, in step 4-2, the feed connection node week load curve, typical building week load curve and Phase

The input variable that weather information is network is answered, the quantity per quasi-representative building load is output variable.

Preferably, in step 4-3, the error function is reflection neutral net desired output and calculates between output By mistake

The function of difference size, kth time sample computing formula is as follows：

Wherein：d_oK () is the desired output of o nodes, yo_oK () is that o nodes calculate output valve, q is output layer god Jing units number；

The partial derivative of each neuron of the calculation error function pair output layer, formula is as follows：

yo_o(k)=f (yi_o(k))

Calculation error is to the partial derivative of each neuron of hidden layer：

Wherein, w_hoFor hidden layer and the connection weight of output layer, yi_oIt is the input vector of output layer o nodes, yi_o(k) It is the input vector of kth time sample output layer o nodes, ho_hK () is the output vector of kth time sample hidden layer h nodes, d_o K () is the desired output of kth time sample o nodes, yo_oK () is the output vector of output layer o nodes, yo'_oK () is defeated Go out the partial derivative of the output vector of layer o nodes to the input vector of o nodes, b_oIt is the threshold value of each neuron of output layer, f X () is activation primitive.

Preferably, in step 4-5, the formula for calculating global error is as follows：

With immediate prior art ratio, the present invention provide technical scheme there is following excellent effect：

Composition and daily electrical characteristics that the present invention passes through collection typical urban building load, are built based on EnergyPlus softwares Different type typical building load model is found, one kind has been proposed on this basis urban distribution network feeder load is separated into into difference The discrimination method of type typical building load, is to specify rational DSM policies and mechanism to lay the foundation.

Description of the drawings

Fig. 1 is that a kind of urban distribution network feeder load based on artificial neural network that the present invention is provided constitutes discrimination method stream Cheng Tu

Specific embodiment

The present invention is described in further detail below in conjunction with the accompanying drawings.

Composition and daily electrical characteristics that this method passes through collection typical urban building load, are built based on EnergyPlus softwares Different type typical building load model is found, one kind has been proposed on this basis urban distribution network feeder load is separated into into difference The discrimination method of type typical building load.It it is critical only that：1) typical load model library should be able to include feeder line institute to be separated In the main loads type in region；2) typical load model can effecting reaction season, date and external environmental factor to electricity consumption The impact of curve.As shown in figure 1, concrete steps include：

Step 1, collect typical building load type (being set to m classes), history in feed connection node correspondence building climate area and use Electrical information and corresponding weather information；

Step 2, the detailed physical model for being set up all kinds of typical buildings respectively based on EnergyPlus softwares are simultaneously simulated all kinds of The actual electricity consumption situation of building load；

Step 3, obtained by the EMS (Energy Management System, EMS) of grid company and treated The all load curves of assessment feeder line, the large scale industry load with station-service EMS can obtain that the load of load can be recognized by station-service EMS Curve, both subtract each other so as to the load curve of part to be identified is obtained.

Step 4, based on load curve to be identified corresponding date and weather information, it is each by EnergyPlus Software Creates The load curve of quasi-representative building load correspondence period；Based on the load curve of above-mentioned typical building load correspondence period, adopt Artificial neural network algorithm carries out the identification of load structure to feeder load to be identified；

Wherein, artificial neural network algorithm uses BP neural network, that is, employing to press Back Propagation Algorithm in step 4 The Multi-layered Feedforward Networks of training, are divided into following steps again.

Step S1：Netinit, to each connection weight initial value, computational accuracy and maximum study number of times are assigned, and arrange k=0；

Step S2：Using all load curves of n quasi-representatives building and corresponding weather information of EnergyPlus Software Creates, profit Sampled to form differently composed feed connection node week load curve with Monte Carlo, and it is (input layer, hidden to form three layers of BP network structures Containing layer and output layer) feedforward link input layer training sample set.Wherein all load curves of feed connection node load, classification is typical All load curves and corresponding weather information of building load is the input variable of network, is defeated per the quantity of quasi-representative building load Go out variable；

Step S3：Using network desired output and reality output, the local derviation of each neuron of calculation error function pair output layer Number, using the inclined of each neuron of the output calculation error function pair hidden layer of connection weight and hidden layer of hidden layer to output layer Derivative；

Wherein Error Calculation function is the function for reflecting neutral net desired output and calculating error size between output, the It is represented by after k time sample is calculated：

Wherein：d_oK () is the desired output of o nodes, yo_oK () is that o nodes calculate output valve, q is output layer god Jing units number.

Error function is to the partial derivative of each neuron of output layer：

yo_o(k)=f (yi_o(k))

Calculation error is to the partial derivative of each neuron of hidden layer：

Wherein, w_hoFor hidden layer and the connection weight of output layer, yi_oIt is the input vector of output layer o nodes, yi_o(k) It is the input vector of kth time sample output layer o nodes, ho_hK () is the output vector of kth time sample hidden layer h nodes, d_o K () is the desired output of kth time sample o nodes, yo_oK () is the output vector of output layer o nodes, yo'_oK () is defeated Go out the partial derivative of the output vector of layer o nodes to the input vector of o nodes,

b_oIt is the threshold value of each neuron of output layer, f (x) is activation primitive.

Step S4：Connection weight is corrected using the output of each neuron of output layer and each neuron of hidden layer, using hidden Input containing each neuron of layer and each neuron of input layer is correcting connection weight；

Step S5：Calculate global error；

Step S6：If study number of times is more than or equal to maximum times m of setting, terminate calculating；If fruit study number of times is less than The maximum times of setting, into step S7；

Step S7：If network error meets required, terminate calculating；If be unsatisfactory for, k=k+1, progressive step S2.

Finally it should be noted that：Above example is most only to illustrate technical scheme rather than a limitation Pipe has been described in detail with reference to above-described embodiment to the present invention, and those of ordinary skill in the art should be understood：Still The specific embodiment of the present invention can be modified or equivalent, and without departing from any of spirit and scope of the invention Modification or equivalent, it all should cover in the middle of scope of the presently claimed invention.

Claims (6)

1. a kind of urban distribution network feeder load based on artificial neural network constitutes discrimination method, it is characterised in that methods described Including：

(1) equipment composition, history power information and the weather information of correspondence climatic province typical building load are collected；

(2) the power load model of all kinds of typical buildings is set up；

(3) based on feeder load curve to be identified corresponding date and weather information, generate all kinds of typical building loads to correspondence when The load curve of section；

(4) identification of load structure is carried out to the feeder load to be identified using BP artificial neural network algorithms.

2. method according to claim 1, it is characterised in that in the step (3), the feeder load curve to be identified is Acquirement is subtracted each other by feeder line to be assessed week load curve and the load curve that can recognize load.

3. method according to claim 1, it is characterised in that in the step (4), comprise the steps：

Step 4-1, netinit, to each connection weight initial value, computational accuracy and maximum study number of times are assigned, and arrange k=0；

Step 4-2, all load curves of n quasi-representatives building for generating and corresponding weather information, using Monte Carlo sampling feedback is constituted The all load curves of line node, and form three layers of BP network structures, including input layer, hidden layer and output layer；

Step 4-3, using network desired output and reality output, the partial derivative of each neuron of calculation error function pair output layer, Connection weight and the output of hidden layer using hidden layer to output layer, the local derviation of each neuron of calculation error function pair hidden layer Number；

Step 4-4, connection weight is corrected with the output of each neuron of output layer and each neuron of hidden layer, it is each using hidden layer The input of neuron and each neuron of input layer is correcting connection weight；

Step 4-5, calculating global error；

If step 4-6, study number of times are more than or equal to maximum times m of setting, terminate calculating；If study number of times is less than setting Maximum times, into step 4-7；

If step 4-7, network error meet required, terminate calculating；If be unsatisfactory for, k=k+1 goes to step 4-2.

4. method according to claim 3, it is characterised in that in step 4-2, the feed connection node week load curve, The typical building week load curve and corresponding weather information are the input variable of network, are per the quantity of quasi-representative building load Output variable.

5. method according to claim 3, it is characterised in that in step 4-3, the error function is reflection nerve net The function of error size between network desired output and calculating output, kth time sample computing formula is as follows：

$e=\frac{1}{2}{\mathrm{\Σ}}_{o=1}^q{\left(d_o\right(k)-{\mathrm{yo}}_o(k\left)\right)}^2$

Wherein：d_oK () is the desired output of o nodes, yo_oK () is that o nodes calculate output valve, q is output layer neuron Number；

The partial derivative of each neuron of the calculation error function pair output layer, formula is as follows：

$\frac{\∂e}{\∂w_{ho}}=\frac{\∂e}{\∂{\mathrm{yi}}_o}\frac{\∂{\mathrm{yi}}_o}{\∂w_{ho}}$

$\frac{\∂{\mathrm{yi}}_o\left(k\right)}{\∂w_{ho}}=\frac{\∂\left({\Σ}_h^p{w}_{ho}{\mathrm{ho}}_h\right(k)-b_o)}{\∂w_{ho}}={\mathrm{ho}}_h\left(k\right)$

$\begin{array}{c}\frac{\∂e}{\∂{\mathrm{yi}}_o\left(k\right)}=\frac{\∂\left(\frac{1}{2}{\Σ}_{o=1}^q{\left(d_o\left(k\right)-{\mathrm{yo}}_o\left(k\right)\right)}^2\right)}{\∂{\mathrm{yi}}_o\left(k\right)}\\ =-\left(d_o\right(k)-{\mathrm{yo}}_o(k\left)\right){\mathrm{yo}}_o^{\′}\left(k\right)=-\left(d_o\right(k)-{\mathrm{yo}}_o(k\left)\right)f^{\′}\left({\mathrm{yi}}_o\right(k\left)\right)\end{array}$

yo_o(k)=f (yi_o(k))

Calculation error is to the partial derivative of each neuron of hidden layer：

$\frac{\∂e}{\∂w_{ho}}=\frac{\∂e}{\∂{\mathrm{yi}}_o}\frac{\∂{\mathrm{yi}}_o}{\∂w_{ho}}=-{\mathrm{\δ}}_o\left(k\right){\mathrm{ho}}_h\left(k\right)$

$\frac{\∂e}{\∂w_{ih}}=\frac{\∂e}{\∂{\mathrm{hi}}_h\left(k\right)}\frac{\∂{\mathrm{hi}}_h\left(k\right)}{\∂w_{ih}}$

$\frac{\∂{\mathrm{hi}}_h\left(k\right)}{\∂w_{ih}}=\frac{\∂\left({\mathrm{\Σ}}_{i=1}^n{w}_{ih}{x}_i\right(k)-b_h)}{\∂w_{ih}}=x_i\left(k\right)$

$\begin{array}{c}\frac{\∂e}{\∂{\mathrm{hi}}_h\left(k\right)}=\frac{\∂\left(\frac{1}{2}{\Σ}_{o=1}^q{\left(d_o\left(k\right)-{\mathrm{yo}}_o\left(k\right)\right)}^2\right)}{\∂{\mathrm{ho}}_h\left(k\right)}\frac{\∂{\mathrm{ho}}_h\left(k\right)}{\∂{\mathrm{hi}}_h\left(k\right)}\\ =-\left({\Σ}_{o=1}^q{\mathrm{\δ}}_o\right(k\left)w_{ho}\right)f^{\′}\left({\mathrm{hi}}_h\right(k\left)\right)-{\mathrm{\δ}}_h\left(k\right)\end{array}$

Wherein, w_hoFor hidden layer and the connection weight of output layer, yi_oIt is the input vector of output layer o nodes, yi_oK () is kth The input vector of secondary sample output layer o nodes, ho_hK () is the output vector of kth time sample hidden layer h nodes, d_oK () is The desired output of kth time sample o nodes, yo_oK () is the output vector of output layer o nodes, yo'_oK () is output layer Partial derivative of the output vector of o nodes to the input vector of o nodes, b_oIt is the threshold value of each neuron of output layer, f (x) is sharp Function living.

6. method according to claim 3, it is characterised in that in step 4-5, the formula for calculating global error is such as Under：

$E=\frac{1}{2m}{\Σ}_{k=1}^m{\Σ}_{o=1}^q{\left(d_o\right(k)-y_o(k\left)\right)}^2.$