Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 10

Electricity load forecasting (Group course work):

objective of this assignment is to investigate the development of 1-day ahead electricity load
forecasting models for the Northern Ireland power system using linear (ADALINE) models and non-
linear Multilayer Perceptron (MLP)

ubmit an individual 1-page (max 500 words) collaboration report containing:


• A statement of their responsibilities in the team and how they contributed to the assignment.
• A description of the responsibilities and contributions of the other team members.
• Details of how the team collaborated (e.g. meetings online/in person, tools used, etc.).
• A reflection on how well the team performed – Did you work well as a team? Did everyone
contribute equally? What were the strengths and weakness of the team? What issues did you
For the assignment you must use the data from 2014, 2015 and 2016 as
training and validation data for model development
Abstract:
The purpose of this report is to research the generation of   linear (ADALINE) models
for 1-day ahead electricity load forecasting for the Northern Ireland electricity
network.
The following report introduces the ANN and then the variables for estimation of the
consumption values are introduced. Although ANNs are usually used to predict short-
term consumptions, this study demonstrates that electricity forecasting at Northern
ireland an artificial neural network power sector could done with greater accuracy.
ANNs are essentially just mathematical tools created to carry out a wide range of
tasks. ANNs consist of an inter-connection of a number of neurons.
ADALINE (Adaptive Linear Neuron or later Adaptive Linear Element) is an early
single-layer artificial neural network and the name of the physical device that
implemented this network. Adaline is a multi-node, single-layer neural network where
each node accepts multiple inputs and produces one output. Training is a process by
which the network learns to recognize the present pattern from the input data set.

In order to learn without a teacher, neural networks need some self-organization


property to keep track of previously seen pattens and answer on the basis of some
familiarity criterion to the current input. Input patterns are grouped into clusters on
the basis of each other’s similarity. The processing elements each have a number of
internal parameters is called weights. This initial set of weights is considered as the
initial conditions for the run.
Introduction:
In practice, the ideal model is one that performs the best when making accurate
predictions for the given data. The ANNs consist of an interconnection of several
neurons. In point of fact, the majority of industrial processes are highly nonlinear.
Due to their strong nonlinear fitting capability, artificial neural networks (ANNs) are
frequently utilized in electricity forecasting prediction. In evaluating models, the terms
"training dataset," "validation dataset," and "test dataset" are clearly defined.
Adaptive linear neuron is described in given below figure. Labels is a cell array with a
brief description of each of the variables in the Xi matrices, while Datei is a vector
that contains the date-time information for the values in Xi.

In practical terms, many variables cannot be measured at all or frequently enough


with traditional instrumentation, owing to the technological or economical constraints
of insufficient space, poor environmental conditions, extreme operating conditions,
etc. In Adraline, hidden nodes with appropriate nonlinear transfer functions are used
to process the information received by the input nodes. So, we use variable selection
method in this groupwork.
A variable selection method known as forward stepwise selection (or forward
selection):
1)Starts with a model that doesn't have any variables—the Null Model.
2) Then starts adding the most significant variables one after the other
3)Until the model incorporates all of the variables under consideration or until a
predetermined stopping rule is reached.

Other input variables can be identified and inserted into the model but if we
specifically take only 10 variables. Our algorithm uses weather information for
modelling.

Artificial neural Variable selection


Null model
networks -Adaline model

Adding significant Process repeats


variables untill combining all
variables- model
completes
Generation or collection of labelled data
cleaning of data on temperature select algorithm

test set
training set validation set

training
final model prediction for given data ahead.

TABLE:
The standard deviation of the residuals is defined as the Root Mean Square Error
(RMSE) (prediction errors). Residuals are a measure of how far away data points are
from the regression line; RMSE is a measure of how spread out these residuals are.
In other words, it indicates how concentrated the data is around the best fit line. To
validate experimental results, root mean square error is commonly used in
climatology, forecasting, and regression analysis.
A table detailing the Forward Selection model variables and the training, validation
and test RMSEs for each dataset. The training is fairly independent of the initial
values of the weight space.

 Training Dataset: The sample of data used to fit the model. The goal here is
to choose the weights of the network to achieve a desired input or output
relationship. This process is known as training the network. The training time
required is dictated by various elements including the complexity of the
problem, the number of data, the structure of network and the training
parameters used.

 Validation Dataset: One of the common techniques to avoid overfitting in


neural networks is to employ a small set of independent data called validation
set and can be combined with the training set. During the learning procedure,
the error shows first a decrease on the validation set as well as on the training
set on the basis of underlying input or output function. After some training
steps u*, the network starts overfitting the spurious and misleading features of
the training set. It is better to stop the learning procedure at this step u* with
smallest error on the validation set as network gives best generalization
performance. This technique is called as early stopping as shown in figure 10
below.
The main goal of a neural network is thus generalization of prediction of given data
of the relationship derived from the training set.

 Test Dataset: The collection of data used to provide an impartial assessment


of a final model fit on the training dataset.

a script to calculate the correlation and plot the relationship between each candidate


predictor field and also we refer to ExploreData.m for code useful for this
task). Therefore, examine which variables are more promising as predictors
of bias and which are less promising.

Table shows the 10 most promising variables with their correlation values and some
summary statistics. Table gives us the indication that the model has better
generalization results and prediction accuracy. With the finalized model, we can save
the model for later or operational use.

DISCUSSION:

In recent years, the accuracy of SONI's load forecasting methods has


deteriorated due to changes in consumer consumption patterns and the impact
of small-scale (unmetered) distributed generation on the grid. This is the one of the
best studies for presentation of an algorithm based on ANN is to predict electricity
consumption in one day ahead. In practical terms, having too many nodes
is not desirable, but you must have enough for the network to grasp the complexity
of the input-output relationships. If our network has only one output node and
we think that the required input-output relationship is fairly simple, so we start with
a hidden layer dimensionality that is two-thirds the input dimension.
Network dimension:

Use the ANN method to estimate the relationship between


inputs and outputs. One may ask what is the smallest number of hidden layers and
units per layer necessary to approximate a particular set of data with a forecast
prediction. Answer is the number of layers it was proven that with at most two hidden
layers with sufficient number of variables it is possible to approximate any function. If
the function is continuous, only one hidden layer is sufficient.

 To find the right number


hidden nodes, repeat these steps with different architectures
and training parameters for networks with a
to q nodes in your hidden layer. Pruning is the removal of neural connections that
have no significant impact on the functionality of the Network.

why you think these variables affect the demand for electricity?

A list of variables that are not useful to consider and why?

Observation:
This work attempts to identify the most suitable network for the desired model after
the characteristics of the ANN.ADALINE (Adaptive Linear Neuron) is easier to
implement, require much fewer computing resources and much faster to execute.
However, it has a drawback which is explained in sequence below. Trial and error is
all well and good in impending the model, but we will need some sort of reasonable
starting point.

Any algorithm that employs the problem-solving heuristic of making the locally
optimal choice at each stage is referred to as a greedy algorithm. In many cases, a
greedy strategy does not result in an optimal solution; however, a greedy heuristic
can produce locally optimal solutions that are approximately equivalent to a globally
optimal solution within a reasonable amount of time. Clearly, forward selection model
is a greedy algorithm.

a plot comparing the predicted and an actual load over the test dataset (i.e., with
time on the x-axis and both the actual and predicted load plotted on y-axis) for each
of the datasets. Comment on the patterns you observe and what they mean

Purpose of neural networks:

Neural networks have helped in following fields of knowledge: machine vision,


speech recognition, robotics control and expert systems. The effectiveness of a
neural network algorithm strongly depends on hardware that executes it. One of the
most interesting aspects of neural networks is their learning capability. Neural
network learning is done by adoptively changing the interconnection strengths
between the neurons. Pruning is the process of removing neurons that do not
significantly contribute to the performance of the network. Pruning can also be
implemented in the ANN algorithm.
To evaluate the resulting ANN’s performance, the following percentage error is used
throughout this report. The difference between the prediction and the actual load is
usually considered as a stochastic process.

actual load−forecast load


Error ¿ ∗100
actual load
An output vector is produced and target outputs, the network’s weights {w ij} are
adjusted to reduce the output error. The error at the output layer propagates
backward to the hidden layer until it reaches the input layer. Once the neural network
is trained, it produces very fast output for a given input data.
In order to forecast future load from the trained ANN, we need to use the recent load
and temperature data in addition to the predicted future temperature. So we predict
future load by inserting the predicted weather information into the predetermined
functional relationship.
Contributions:

Your report should contain the following:


• Descriptions of the models investigated, the procedures used to obtain your
results and explanations and observations on the results obtained.
• Any plots included should be fully annotated (i.e. appropriate title, legend
and axis labels, etc.) and appropriately scaled so important details are clear.
• The names of the m-files you used to get the results presented.

Conclusion:
In applied machine learning, the meticulous design of our test harness is
absolutely necessary. This report deals with the most popular ANN learning
algorithms based on competitive learning for electricity forecasting. When
making purchases of electricity through the Integrated Single Electricity
Market (I-SEM), the System Operator for Northern Ireland (SONI) makes use
of short-term forecasts for the day ahead and during the day using ANN.ANN
paradigms are all roughly based on the simulation of some specific features of
biological nervous systems. ANN’s decision processing remains quite opaque
and a translation into meaningful symbolic knowledge is difficult to perform.
Finally cross, auto and backward connections transform ANN structure from
static to dynamic allowing the recognition and production of time sequences
with different dynamics approach.
Information from the incoming nodes is processed by using hidden nodes with
suitable nonlinear transfer functions. The terms "validation" and "test" sets
frequently have their meanings reversed in the machine learning literature.
Finally, we will be able to rely even more on the estimated performance with a
test harness that is stronger. ANN learn by example and cannot be
programmed to perform a specific task. Compare to other methods ANN
allows more flexible relationship between temperature and load pattern. In
general, trained ANN requires only 4 to 20 milliseconds for testing.

References:

You might also like