Sie sind auf Seite 1von 8

PREDICTION INTERVAL

ESTIMATION OF
ELECTRICITY PRICES
USING NEURAL
NETWORKS
Submitted by : Masiha Aijaz
M. Tech EPSM 1st sem
R.no : 10

Introduction

Electricity price forecasting is essential for market participants in both daily operation and longterm planning analyses, such as designing bidding strategies and making investment decisions.
Because the electricity price is a stochastic time series with very high uncertainty , it is very
difficult to predict the exact value of the future price. Consequently, in addition to predicting the
price value, predicting the distribution of the future price, i.e., the prediction interval becomes
highly meaningful.
Price intervals can effectively reflect the uncertainties in the predication results. Generally
speaking, a prediction interval is a stochastic interval, which contains the true value of the price
with a preassigned probability. Because the prediction interval can quantify the uncertainty of the
forecasted price, it can be employed to evaluate the risks of the decisions made by market
participants.
There are two major challenges for accurately electricity price interval forecasting of the
electricity price:
1) To estimate the prediction interval, the value of the future price should be accurately
forecasted. However, this is difficult because the electricity price is a nonlinear time series,
which is highly volatile and cannot be properly modeled by traditional linear time series models.
2) The variance of the price should also be accurately forecasted. This is because it is essential to
estimate the price distribution so as to estimate the prediction interval
We consider the task of predicting the future electricity load (demand) from a time series of
previous electricity loads. In particular, we study a time series of electricity loads measured every
30 minutes and our goal is to make predictions for the next value. Forecasting with lead times up
to an hour is classified as Very Short Term Load Forecasting (VSTLF) and is a key task in the
operation of electricity markets.
The electricity load series data is complex. It is highly non-linear and non-stationary, with
several cycles that are nested into each other, e.g. daily, weekly and yearly. In addition, it
includes random components resulting from large industrial loads with irregular hours of
operation, special events, holidays, extreme weather and sudden weather change. It is also noisy
due to malfunctioning of data collection devices. All these factors combined together make the
task of building accurate prediction models challenging.

In this review, focus is on interval forecasting , i.e. at time t the task is to predict an interval of
values for time t+h with a certain probability, called confidence level. More specifically, an
interval forecast, or Prediction Interval (PI), consists of an upper and lower bound, between
which the future value is expected to lie with a certain prescribed probability. Predicting an
interval of values with a certain probability instead of just predicting a single value gives more
information about the variability of the target variable and the associated uncertainty.
They are especially suitable for applications requiring balancing of demand and supply, e.g.
electricity and financial markets, stock manufacturing and inventory management
A method for constructing PIs using NNs is reviewed .It is called LUBE and its main idea is to
directly generate the lower and upper bounds of PIs using a feedforward NN with two outputs.
LUBE uses the point forecasts for the training data and a specially designed cost function that is
minimized in order to generate high quality PIs for new data.

PROBLEM STATEMENT
Given is a time series containing n observations
In a standard point forecasting scenario, the goal is to forecast
steps ahead, using the data up to time n .

,the value of the series h

In an interval prediction scenario, our goal is to forecast the PI of


upper bounds,

i.e. its lower and

In this paper we consider a forecasting horizon h = 1, so our task is to predict


, the PI for a 1-step ahead point forecast.
Given a confidence level
condition :
contain

and

a valid PI for
will satisfy the following
i.e. the probability that the interval will

will be equal or greater than the prescribed confidence level.

EVALUATING THE QUALITY OF PI


The two most important properties of PI are coverage probability and width. A good PI will
have high coverage probability and small width.

Prediction Interval Coverage Probability (PICP)


Given a dataset of N examples, it is the probability that the target value
example will fall between the upper limit
interval

and lower limit

of the i -th

of the prediction

, averaged over all i.

A required nominal confidence level


is set in advance. PIs with PICP
smaller than
are not reliable. PICP alone is not enough to measure the quality of a PI.
The width of the interval is related to its coverage probability wider PIs have high
coverage probability and vice versa. Thus, it is possible to achieve the required minimum
PICP by simply widening the interval. However, too wide intervals are useless in
practice.

Coverage Width-based Criterion (CWC)

The two main principles of CWC are:


1) If the coverage probability is above the confidence threshold, i.e.
then CWC should depend only on the PIs width.
2) If the coverage probability is below the confidence threshold, i.e. the PIs are not
reliable

The LUBE Method


The LUBE method computes PIs using feedforward NNs. It was shown to produce high quality
PIs at a reasonable computational cost.
The main idea is to train a NN to predict the PIs. The network learns to compute PIs on a training
set and is then used to predict PIs on a new set of examples.
The architecture of the LUBE NN is a standard multi-layer perceptron as shown in Fig.1. There
are p input neurons, corresponding to the input variables of each example, and two output
neurons corresponding to the lower and upper bounds of the PI for this example. There are one
or more hidden layers and the number of neurons in them is set experimentally

The LUBE NN is trained to minimize the CWC cost function. As CWC is not differentiable, the
backpropagation algorithm cannot be used to do this. Instead, the LUBE NN uses the simulated
annealing algorithm

After initialization of the parameters, including the starting temperature To

which is set to a high value, the starting solution (NN weights) w0 and the starting cost function
CWCo, the NN is used to learn PIs for the training data. The two target values
of the prediction interval

are both set to the target point forecast

The cost function CWCnew is then computed and minimized. If there is an improvement, i.e.
CWCnew is less than the previous best CWC, CWCnew are accepted.
The acceptance depends on the badness of the new solution. Hence, bad solutions are more
likely to be accepted at the beginning of the training. If the stopping criterion is not satisfied, a
new set of weights is obtained via small perturbation and the training continues.
When the training is completed, the trained NN is used to compute the PIs for the testing data.

E XPERIMENTAL SETUP
Data
The electricity load data for the state of New South Wales in Australia for the year 2010 is
selected. The data is provided by the Australian Energy Market Operator (AEMO) and is
publicly available. The data is sampled every 30 minutes and recorded for each month. Thus
there are 12 case studies in total, one for each month, and each of them contains between 1344
samples (for February) and 1488 samples (for the 31 days long months).
For each case study, the available data set is divided into three non-overlapped subsets: training
D train, validation D valid and testing D test. The split was 50%-30%-20% respectively.
The training set is used for feature selection and training of the NN models, the validation set is
used to select the best NN architecture and the testing set is used to evaluate the performance of
the proposed method for PI construction.

Quality of PIs for Testing Data


To evaluate the quality of the constructed PIs for the testing data, we use the two performance
indices introduced PICP and CWC. The results are shown in Table II.

Table shows that the coverage probability PICP is higher than the nominal confidence level of
90%, for all 12 case studies. This means that the constructed PIs are valid.The mean PICP and
standard deviation over all case studies were 97.44 3.29 indicating that in most of the case
studies, the constructed PIs considerably outperformed the prescribed confidence level e.g. in 7
cases PICP was above 99% and only in 3 cases it was between 91.4% and 93.3%
In the CWC cost function we used =50 and nominal confidence level =0.9. High value of
will highly penalize the PIs with a coverage probability below the nominal confidence level.

CONCLUSION
The task of interval forecasting for electricity load prediction. We adopted and extended
theLUBE method, a NN-based method for estimating the lowerand upper bound of PIs. The
extended method, called We conducted a comprehensive evaluation using Australian electricity
load data for one year, sampled every 30 minutes. The results showed that LUBEX was able to
produce high quality PIs i.e. PIs with a high coverage probability (97.44%) that satisfy the
minimum confidence level, and were very stable over multiple runs
A comprehensive evaluation using Australian electricity load data for one year, sampled every
30 minutes. The results showed that LUBEX was able to produce high quality PIs i.e. PIs with
a high coverage probability (97.44%) that satisfy the minimum confidence level, and were very
stable over multiple runs.

REFERENCES
I. Koprinska, M. Rana, and V. G. Agelidis, "Yearly and seasonal
models for electricity load forecasting," in Proc. International Joint
Conference on Neural Networks (IJCNN) , San Jose, 2011, pp. 1474-1481.

Das könnte Ihnen auch gefallen