CN106845705A - The Echo State Networks load forecasting model of subway power supply load prediction system - Google Patents

Abstract

A kind of electric power supply system for subway Short Term Load Forecasting System, including load statistical module, load data calling module, load data prediction module, predicated error statistical module, images outputting module and data outputting module；Wherein load statistical module is used for statistical history load data, load data calling module calls the historical load data in load statistical module and is sent to load data prediction module, load data prediction module is predicted and exports prediction data according to historical load data to following load, after predicated error statistical module is calibrated to the prediction data for exporting, the data of revised prediction are exported by images outputting module and data outputting module；Wherein described load data prediction module is built using the electric power supply system for subway Short-term Load Forecasting Model based on Echo State Networks.Forecast model based on Echo State Networks has good precision of prediction and prediction stability.

Description

The Echo State Networks load forecasting model of subway power supply load prediction system

Technical field

Echo state nerve net is based on the present invention relates to electric power supply system for subway short-term load forecasting field, more particularly to one kind The electric power supply system for subway Short Term Load Forecasting System of network load forecasting model.

Background technology

City underground traffic is urban modernization mark, be in modern city preferably, citizens' activities demand can be met A kind of public transport.It can also be negatively affected, subway while bringing quick and easy to us to power network Used as high-power nonlinear-load, influence the characteristics of its operation and after putting into operation to area power grid should be that power department is paid close attention to Problem.To estimating more serious problem, the appropriate precautionary measures should be taken, make its negative effect control in allowed band It is interior.The characteristics of electric power supply system for subway load has randomness, fluctuation and unstability, these features cause electric power supply system for subway Load on Electric Power Grid produces impact, disturbs the safe and stable and economical operation of power network.It is pre- by carrying out to electric power supply system for subway load Survey can effectively reduce the adverse effect produced to power network, to ensure that power system can run with security and stability.

Echo State Networks are made up of input layer, reserve pool, output layer, and reserve pool is dynamic network, by largely with The neuron of machine partially connected is constituted.The application of reserve pool overcomes traditional neural network convergence rate slowly and is easily absorbed in local pole Small problem.Echo State Networks have been applied to the multiple research field such as photovoltaic power generation power prediction.But subway power supply System loading forecasting research is a current research blank, in particular by Echo State Networks to electric power supply system for subway Load is predicted in research vacuum.

The content of the invention

In order to fill up the blank of prior art, the present invention proposes a kind of electric power supply system for subway Short Term Load Forecasting System, Including load statistical module, load data calling module, load data prediction module, predicated error statistical module, images outputting Module and data outputting module；Wherein load statistical module is used for statistical history load data, and load data calling module is adjusted With the historical load data in load statistical module and load data prediction module is sent to, load data prediction module is according to going through History load data is predicted and exports prediction data to following load, prediction data of the predicated error statistical module to output After being calibrated, the data of revised prediction are exported by images outputting module and data outputting module；It is wherein described negative Lotus data prediction module is built using the electric power supply system for subway Short-term Load Forecasting Model based on Echo State Networks 's.

Further, under the electric power supply system for subway Short-term Load Forecasting Model based on Echo State Networks is used State algorithm acquisition：Build Echo State Networks, including input layer, reserve pool, output layer；Input layer has k in network model Individual input node, reserve pool has n internal node, and output layer has l output node；

In t, the input vector of network is u (t)=[u₁(t),u₂(t),···,u_k(t)]^T, internal state to It is x (t)=[x to measure₁(t),x₂(t),···,x_n(t)]^T, output vector is y (t)=[y₁(t),y₂(t),···,y_l (t)]^T, then the state equation and output equation of ESN forecast models be respectively：

X (t+1)=f (wⁱⁿu(t+1)+wx(t)+w^backy(t)) (1)

Y (t+1)=f^out(w^out(u(t+1),x(t+1),y(t+1))) (2)

In formula, wⁱⁿInput layer to the input connection weight matrix of reserve pool is represented, w represents reserve pool inside connection weight square Battle array, w^backRepresent output layer to the feedback link weight matrix of reserve pool, w^outRepresent reserve pool to the output connection weight of output layer Value matrix, f represents the excitation function of deposit pool unit, takes hyperbolic tangent function；f^outThe excitation function of output unit is represented, is taken Identity function；

Wherein output connection weight matrix w^outBy give training sample (u (t), y (t), (t=1,2, Q)) determine, its training process can be divided into two stages：

(1) sample phase

Original state first to network carries out assignment, and the original state of normal conditions lower network is 0, i.e. x (t)=0.So Afterwards by training sample (u (t), t=1,2, q) pass through wⁱⁿBe input in dynamic reserve pool, according to formula (1) and formula (2) according to The secondary calculating for completing network state vector x (t) and output vector y (t)；

From a certain moment h start recordings Echo State Networks internal system state variable and corresponding sample data, Then phasor ([u is used₁(j),u₂(j),···,u_k(j)]^T；[x₁(j),x₂(j),···,x_n(j)]^T) constitute matrix B (q-h+1), and with phasor ([y₁(j),y₂(j),···,y₃(j)]^T) come constitute matrix T (q-h+1, l).Wherein j=h, h+ 1,···,q；

(2) the weight computing stage

According to internal state data and sample data that Echo State Networks system in sampling process is recorded, by most Small square law linear regression is calculated output connection weight matrix w^out.Due to state variable x (t) and net of the inside of network The reality output of networkBetween be linear relationship, so using the reality output of networkCarry out the preferable output y of Approximation Network (t)：

Calculating weight w_i ^outProcess need to ensure that the root-mean-square error of above-mentioned formula is minimum that then problem can be converted To solve the optimization problem of formula below：

From from the perspective of mathematics, this is a problem for linear regression, can be converted to and ask the inverse matrix of matrix B to ask Topic, i.e. w^out=B^-1T, Echo State Networks training terminates.

The present invention devise statistics with prediction one electric power supply system for subway Short Term Load Forecasting System, it is proposed that based on return The electric power supply system for subway Short-term Load Forecasting Model and its construction method of sound state neutral net.Echo State Networks are by defeated Enter layer, reserve pool, output layer to constitute, reserve pool is dynamic network, and the neuron connected by a large amount of Random sparseness is constituted.Reserve pool Application to overcome traditional neural network convergence rate slow and be easily absorbed in the problem of local minimum.Using actual electric power supply system for subway Historical data carry out simulating, verifying, it is good pre- that simulation result shows that the forecast model based on Echo State Networks has Survey precision and prediction stability.

Brief description of the drawings

Fig. 1 is the schematic diagram of electric power supply system for subway Short Term Load Forecasting System of the invention；

Fig. 2 is the Echo State Networks forecast model of electric power supply system for subway Short Term Load Forecasting System of the invention Schematic diagram；

Fig. 3 is that electric power supply system for subway Short Term Load Forecasting System of the invention compares bent with the prediction of BP-NN forecast models Line chart.

Specific embodiment

In comprehensively deep electric power supply system for subway Load Characteristic Analysis and the base of accurate advanced load forecasting model research On plinth, the electric power supply system for subway Short Term Load Forecasting System of statistics and prediction one is devised.As shown in figure 1, of the inventionly Iron electric power system Short Term Load Forecasting System include load statistical module, load data calling module, load data prediction module, Predicated error statistical module, images outputting module and data outputting module.Wherein load statistical module is negative for statistical history Lotus data, load data calling module calls the statistics in load statistical module and is sent to load data prediction module, Load data prediction module is predicted and exports prediction data according to historical load data to following load, predicated error system After meter module is calibrated to the prediction data for exporting, export revised by images outputting module and data outputting module The data of prediction.

In the electric power supply system for subway Short Term Load Forecasting System, load prediction module is nucleus module, based on echo The electric power supply system for subway Short-term Load Forecasting Model of state neutral net is a kind of core algorithm of the module.

Echo State Networks are a kind of New Recursive neutral nets, and its network structure is by input layer, reserve pool, defeated Go out layer composition.Reserve pool is a dynamic network, and it is that the neuron connected by a large amount of Random sparseness is constituted, and works as input signal During into reserve pool inside, its internal complicated non-linear state space can be excited, network signal is then exported by output layer. Conventional recursive neural network BP training algorithm is complicated, computationally intensive, and deposit is set up in Echo State Networks forecast model Pond and completion network training are carried out respectively, only need to adjust reserve pool to the weights of output layer in network training, other Weights just no longer change after netinit, and training algorithm is simple, and amount of calculation is small, can effectively solve the problems, such as local optimum, return Sound state neural network prediction model is as shown in Figure 2.

Solid line connection represents the necessary connection weight of forecast model in Fig. 2, and dotted line connection is not for constituting forecast model It is required, it is the connection weight according to different situations come selection.As shown in Figure 2, input layer has k input section in network model Point, reserve pool has n internal node, and output layer has l output node.In t, the input vector of network for u (t)= [u₁(t),u₂(t),···,u_k(t)]^T, internal state vector is x (t)=[x₁(t),x₂(t),···,x_n(t)]^T, output Vector is y (t)=[y₁(t),y₂(t),···,y_l(t)]^T.Then the state equation of ESN forecast models and output equation are distinguished For：

X (t+1)=f (wⁱⁿu(t+1)+wx(t)+w^backy(t)) (1)

Y (t+1)=f^out(w^out(u(t+1),x(t+1),y(t+1))) (2)

In formula, wⁱⁿInput layer to the input connection weight matrix of reserve pool is represented, w represents reserve pool inside connection weight square Battle array, w^backRepresent output layer to the feedback link weight matrix of reserve pool, w^outRepresent reserve pool to the output connection weight of output layer Value matrix.F represents the excitation function of deposit pool unit, typically takes hyperbolic tangent function；f^outRepresent the excitation letter of output unit Number, typically takes identity function.

Echo State Networks forecast model only need to by give training sample (u (t), y (t), (t=1, 2, q)) determine network output connection weight matrix w^out, its training process can be divided into two stages：Sampling rank Section and weight computing stage.

(1) sample phase

Original state first to network carries out assignment, and the original state of normal conditions lower network is 0, i.e. x (t)=0.So Afterwards by training sample (u (t), t=1,2, q) pass through wⁱⁿBe input in dynamic reserve pool, according to formula (1) and formula (2) according to The secondary calculating for completing network state vector x (t) and output vector y (t).

By being calculated output connection weight matrix w^out, Echo State Networks system need opened from a certain moment h Beginning records its internal state variable and corresponding sample data, then with phasor ([u₁(j),u₂(j),···,u_k(j)]^T；[x₁ (j),x₂(j),···,x_n(j)]^T) constitute matrix B (q-h+1), and with phasor ([y₁(j),y₂(j),···,y₃(j) ]^T) come constitute matrix T (q-h+1, l).Wherein j=h, h+1, q.

(2) the weight computing stage

According to internal state data and sample data that Echo State Networks system in sampling process is recorded, by most Small square law linear regression is calculated output connection weight matrix w^out.Due to state variable x (t) and net of the inside of network The reality output of networkBetween be linear relationship, so using the reality output of networkCarry out the preferable output y of Approximation Network (t)：

Calculating weight w_i ^outProcess need to ensure that the root-mean-square error of above-mentioned formula is minimum that then problem can be converted To solve the optimization problem of formula below：

From from the perspective of mathematics, this is a problem for linear regression, can be converted to and ask the inverse matrix of matrix B to ask Topic, i.e. w^out=B^-1T, Echo State Networks training terminates.

Using certain city underground electric power system as research object, the historical load data to the electric power supply system for subway carries out phase The simulation analysis of pass, by the Load Characteristic Analysis to electric power supply system for subway, day, type factor was to electric power supply system for subway The influence of load is than larger.Therefore this patent forecast model use input quantity include prediction day day type and prediction a few days ago One day load of synchronization, and first three hour and latter three hours load, and prediction the last week day synchronization is negative Lotus.The output quantity of forecast model is to predict the load value of prediction time day.

Because the presence for having singular value in sample data, the dimension of variable are also different, essence is entered in the prediction for forecast model Degree produces influence, therefore needs to be normalized initial data before network training.Day categorical data is divided three classes, Monday takes 1, and Tuesday to Friday takes 0.2, and Saturday and Sunday take 0.1.

Echo State Networks forecast model and BP forecast models are built respectively to the load of electric power supply system for subway one day Simulation and prediction, simulation result are carried out as shown in figure 3, and predicted value and actual value are compared with the predicted value of BP forecast models Compared with its application condition result is as shown in table 1.

1 two kinds of model predictive errors of table compare

From Fig. 1 and Biao 1 as can be seen that the average forecasting error of Echo State Networks forecast model is predicted than BP-NN The average forecasting error of model improves 2.65%, and the largest prediction error of Echo State Networks forecast model compares BP-NN The largest prediction error of forecast model improves 2.90%, and the precision of prediction for indicating ESN forecast models is pre- apparently higher than BP-NN Survey the precision of prediction of model.

Although be described in detail to the present invention in conjunction with the embodiments, it should be understood by those skilled in the art that Ground is that the various amendments, deformation under without departing substantially from spirit of the invention and essence are all allowed, and they both fall within right of the present invention It is required that protection domain among.

Claims (2)

1. a kind of electric power supply system for subway Short Term Load Forecasting System, it is characterised in that including load statistical module, load data is adjusted With module, load data prediction module, predicated error statistical module, images outputting module and data outputting module；Wherein bear Lotus statistical module is used for statistical history load data, and load data calling module calls the historical load number in load statistical module According to and be sent to load data prediction module, load data prediction module carries out pre- according to historical load data to following load Prediction data is surveyed and exports, after predicated error statistical module is calibrated to the prediction data for exporting, by images outputting module And data outputting module exports the data of revised prediction；Wherein described load data prediction module is using based on echo The electric power supply system for subway Short-term Load Forecasting Model of state neutral net and build.

2. electric power supply system for subway Short Term Load Forecasting System according to claim 1, it is characterised in that described based on echo The electric power supply system for subway Short-term Load Forecasting Model of state neutral net is obtained using following algorithms：

Build Echo State Networks, including input layer, reserve pool, output layer；Input layer has k input section in network model Point, reserve pool has n internal node, and output layer has l output node；

In t, the input vector of network is u (t)=[u₁(t),u₂(t),···,u_k(t)]^T, internal state vector is x (t)=[x₁(t),x₂(t),···,x_n(t)]^T, output vector is y (t)=[y₁(t),y₂(t),···,y_l(t)]^T, then The state equation and output equation of the forecast model are respectively：

X (t+1)=f (wⁱⁿu(t+1)+wx(t)+w^backy(t)) (1)

Y (t+1)=f^out(w^out(u(t+1),x(t+1),y(t+1))) (2)

In formula, wⁱⁿInput layer to the input connection weight matrix of reserve pool is represented, w represents reserve pool inside connection weight matrix, w^backRepresent output layer to the feedback link weight matrix of reserve pool, w^outRepresent reserve pool to the output connection weight of output layer Matrix, f represents the excitation function of deposit pool unit, takes hyperbolic tangent function；f^outThe excitation function of output unit is represented, perseverance is taken Deng function；

Wherein output connection weight matrix w^outBy the training sample that gives, (u (t), y (t), (t=1,2, q)) comes It is determined that, its training process can be divided into two stages：

(1) sample phase

Original state first to network carries out assignment, and the original state of normal conditions lower network is 0, i.e. x (t)=0.Then will Training sample (u (t), t=1,2, q) pass through wⁱⁿIt is input in dynamic reserve pool, it is complete successively according to formula (1) and formula (2) Into the calculating of network state vector x (t) and output vector y (t)；

From a certain moment h start recordings Echo State Networks internal system state variable and corresponding sample data, then With phasor ([u₁(j),u₂(j),···,u_k(j)]^T；[x₁(j),x₂(j),···,x_n(j)]^T) constitute matrix B (q-h+ 1), and with phasor ([y₁(j),y₂(j),···,y₃(j)]^T) come constitute matrix T (q-h+1, l).Wherein j=h, h+ 1,···,q；

(2) the weight computing stage

According to internal state data and sample data that Echo State Networks system in sampling process is recorded, by a most young waiter in a wineshop or an inn Multiplication linear regression is calculated output connection weight matrix w^out, due to state variable x (t) and network of the inside of network Reality outputBetween be linear relationship, so using the reality output of networkCarry out preferable output y (t) of Approximation Network：

$y\left(t\right)=\hat{y}\left(t\right)=\underset{i=1}{\overset{k}{\Σ}}{w_i}^{out}{u}_i\left(t\right)+\underset{i=k+1}{\overset{k+n}{\Σ}}{w_i}^{out}{x}_{i-k}\left(t\right)---\left(3\right)$

Calculating weight w_i ^outProcess need to ensure that the root-mean-square error of above-mentioned formula is minimum that then problem can be converted into and ask Solve the optimization problem of formula below：

$\min \frac{1}{q-h+1}\underset{t=h}{\overset{q}{\Σ}}{\left(y\right(t)-\underset{i=1}{\overset{k}{\Σ}}{w_i}^{out}{u}_i(t)-\underset{i=k+1}{\overset{k+n}{\Σ}}{w_i}^{out}{x}_{i-k}(t\left)\right)}^2---\left(4\right)$

From from the perspective of mathematics, this is a problem for linear regression, can be converted to the inverse matrix problem for seeking matrix B, i.e.,：

w^out=B^-1T (5)

W is calculated by formula (5)^out；

Echo State Networks training terminates.