10 II February 2022

https://doi.org/10.22214/ijraset.2022.40467
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue II Feb 2022- Available at www.ijraset.com

Artificial Neural Network Based Load

Forecasting
Shweta Singh
Power System Engineering, Dr. A.P.J Abdul Kalam Technical University

Abstract: In this report our point is to figure the power costs as precisely as conceivable by planning a half and half model and
utilizing different blunder capacities to really take a look at the exactness of our outcome. Before liberation come to presence
years and years back, the electric power enterprises have been overwhelmed by utilities that had full command over movements
of every kind nearby. In any case, after its first endeavor in Latin America, the business has been on the move in many nations
all over the planet. In a liberated market, end-use clients have the decision to choose their power provider.
Keywords: Load Forecasting, ,Load forecasting models, Electricity pricing forecasting, AI Based Load forecasting, Neural
network model

I. INTRODUCTION
For hundred years there was a regulated electric power industry throughout the world. In regulated market in one area only one
company or government agency that produced, transmitted and sold electric power and services. Regulated market governed by
vertically integrated utilities. So there was monopoly (that means consumers had no choice to choose their commodities supplier) of
government companies in regulated electric market. In early 90’s reforms of power market started which included restructuring and
unbundling of electric market. Reforms were based on privatization of electricity market. Chile was the first country in which
restructuring of market started, followed by some Latin American countries and spread across the world. “Forecasting is a tool used
for predicting future demand based on past demand information.” Forecasting is an essential planning tool that assists the decision-
makers and planners to visualize and plan the future of the system according to their necessity. In electricity market, price and load
forecasting are the two major planning tools for generation, transmission and distribution systems. Forecasting is the process of
making statements about events whose actual outcomes (typically) have not yet been observed.

II. ARTIFICIAL NEURAL NETWORK BASED FORECASTING MODELS

It is seen in writing that old style models, i.e., AR, MA, ARMA and ARIMA models, otherwise called time-series or Box-Jenkins
models, are very little fitting for transient value anticipating (STPF) in ebb and flow climate of high non-linearity and various
irregularity. As a rule, they produce estimating blunder (MAPE) over 10%. Additionally, it saw that the old style models
fundamentally require stationarity in time series information. Yet, the time-series value information in genuine are never fixed, the
training to change over these time-series information to roughly fixed time-series information, to get estimated value determining,
adds stationarity. Along these lines, old style determining models sometimes fall short for STPF. Computerized reasoning (AI)
strategies are tracked down performing admirably, all things considered, issues. These strategies remove the covered up or innate
relationship inside the information series, which can't be investigated effectively by straightforward time-series anticipating models.
In the writing AI-based strategy i.e., Artificial Neural Networks (ANN) is broadly utilized for STPF. This strategy gives the most
ideal decisions to anticipating with non-direct conduct of the value information.Because of liberation in the power market, and
expanded rivalry among power age organizations, cost of the power assumes a significant part for these organizations. They need to
create the power on serious costs, with great quality. In this manner the peculiarity's like power age; value variety, load request and
so on are exceptionally unique in nature, which need an extremely exact estimating. Alluding to broad writing accessible on AI-
based value anticipating models, it is induced that ANN is one of the most appropriate method to be used for value guaging.
Consequently, different ANN-based value determining models are created and worked on in this theory work.
Generally utilized actuation (move) capacities are Sigmoid, Gaussian and other non-direct capacities. It is chosen based on kind of
organization structure, though, a feed forward NN can be prepared utilizing tansigmoid, log-sigmoid and straight actuation
capacities.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1071
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue II Feb 2022- Available at www.ijraset.com

These actuation capacities are given as following:

Sigmoid

Tan-sigmoid:

Gaussian:
Linear:
Where c and σ are the middle and width of Gaussian enactment capacity and k is a consistent, characterizing the slant of the line.

A. Learning Algorithms for ANN

A few learning calculations exist in the writing for preparing fake neural organizations and each preparing strategy is the most ideal
for various kinds of issues. In the proposed work different preparation calculations are tried to track down the most ideal preparation
calculation for anticipating the power cost precisely with least MAPE and ready to observe the right connection between input series
and the power cost.
The preparation calculations considered in this paper are:

B. Conjugate angle Calculation with Powell-Beale Restarts

The inquiry course is occasionally reset to the negative of the inclination and it happens when the quantity of emphasess is
equivalent to the quantity of organization boundaries (loads and predispositions). Powell proposed one such technique which
depended on a previous strategy proposed by Beale. In this technique in the event that the symmetry left between the current
inclination and the past slope is next to no the calculations restarts. The disparity (1) is utilized to test for the symmetry.

If this condition is met, the search direction is reset to the gradient's negative value.
Fletcher-Reeves updates with conjugate gradient backpropagation:
In the initial iteration, this training method searches in the steepest direction (i.e., the negative of the gradient).

A linear search strategy is used to determine the best distance to advance along the current search direction.

The next search direction is calculated in such a way that it is conjugate with the previous search directions. The new search
direction is determined by combining the computed steepest drop direction with the previous search direction:
The Fletcher-Reeves method updates the procedure as:

βk is the ratio of the norm squared of the current gradient to the norm squared of the previous gradient.

C. Conjugate Gradient Using Polak-Ribiere Updates

This version of the conjugate gradient algorithm was proposed by Polak and Ribiére.

In this algorithm βk, If updated as indicated in the norm squared of the prior gradient is divided by the inner product of the previous
change in the gradient with the current gradient.

D. Resilient Back Propagation

Sigmoid exchange work is ordinarily utilized in Multilayer feed forward networks in the secret layers, for crushing the limitless
information range into a limited result range.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1072
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue II Feb 2022- Available at www.ijraset.com

The incline of the sigmoid capacities will in general become zero for the huge information. The steepest plummet calculation while
preparing the multi-facet organization can make the size of the angle tiny and hence, it cause little changes in the loads and
predispositions, despite the fact that the loads and inclinations are a long way from their ideal qualities. The tough back proliferation
preparing calculation invalidates the unsafe impact of progress in extent of the incomplete subsidiaries. The heading of the weight
update in this not entirely settled by the indication of the halfway subordinate and not by the greatness of the subsidiary.
The adjustment of size of the weight is resolved utilizing a different update esteem, the loads and inclinations are expanded by
delta_inc if, the presentation work subordinate has same sign as for that load for two successive cycles and the size is diminished by
delta_dec assuming the exhibition work subsidiary has different sign concerning that weight sign from the previous emphasis.
The loads and inclinations are not refreshed for zero subordinates and the adjustment of size turns out to be enormous in the event
that the adjustment of worth of the weight is in similar course for quite some time.

E. Scaled form inclination Algorithm

In this calculation, the upsides of weight and predilection are refreshed using the scaled form angle technique. This computation can
be used to properly prepare networks with no subordinate capacity for their weight, net information, and move capacities. Back-
engineering is used to determine the execution work subordinates for weight and inclination considerations. Using the wavelet
change as a pre-handler. There are two sorts of Wavelet Transform (WT): constant and discrete. In the extent of this theory, we are
worried about the discrete WT simply because we need to manage time series estimated at discrete time focuses .
The power value The series features value spikes, making it difficult to figure out. The wavelet change is a technique that converts a
set of poorly performed value series into a set of constitutive series that has better properties and can thus be predicted more
effectively and precisely. Because of the separating effect of the wavelet change, the series is divided into low recurrence and high
recurrence sections, which are referred to as surmised and detail series, respectively. Persistent wavelet changes (CWT) and
discontinuous wavelet changes (DWT) are the two types of wavelet changes (DWT).

The sign f (x) CWT W(c, d) for a mother wavelet (x) is given by

The wavelet's dispersion is controlled by the scale parameter c, while its centre position is determined by the translation parameter d.
The DWT is more efficient and accurate than the CWT. DWT, on the other hand, is defined as

where T is the length of the sign f (t) , the scaling and interpretation boundaries are elements of the whole number variablesm and
n(a =2m ,b= n2m ) ; t is the discrete time list.

III. PARTICLE SWARM OPTIMIZATION (PSO)

Swarm Intelligence (SI) is a creative, widely distributed intelligent viewpoint for addressing advancement difficulties that was
inspired by the organic models of vertebrates amassing, rushing, and crowding oddities.
Molecule Swarm Optimization (PSO) consolidates amassing practices saw in groups of birds, schools of fish, or bee hives, and,
surprisingly, human social conduct, from which the thought is arisen. PSO is a populace based streamlining apparatus, which could
be executed and applied effectively to tackle different capacity advancement issues, or the issues that can be changed to work
improvement issues [23]. As a calculation, the principle strength of PSO is its quick intermingling, which contrasts well and
numerous worldwide improvement calculations like Genetic Algorithms (GA), Simulated Annealing (SA) and other worldwide
streamlining calculations. While populace based heuristics are all the more exorbitant due to their reliance straightforwardly upon
work esteems rather than subsidiary data, they are anyway helpless to untimely intermingling, which is particularly the situation
when there are numerous choice factors or aspects to be improved. Taking everything into account, the arrangement swarm is
compared to the bird swarm; the birds' movement from one spot to the next is analogous to the advancement of the arrangement
swarm; great data is analogous to the most positive thinker arrangement; and the food asset is analogous to the most confident
person arrangement throughout the course.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1073
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue II Feb 2022- Available at www.ijraset.com

With everyone's help, the most confident person arrangement can be figured out in a molecular swarm advancement calculation.
The molecule with no quality or volume fills in for each individual, and the basic code of behaviour is enforced for each molecule to
demonstrate the complexity of the complete molecule swarm. Particles, which are potential arrangements, fly through the problem
space in PSO by adhering to the most recent optimal particles Every particle keeps track of its directions in the problem space that
are related to the best arrangement (health) that has been achieved thus far. This value is referred to as pbest. Another best worth
followed by the molecule swarm analyzer is the best worth obtained over such a lengthy distance by any molecule in the molecule's
neighbours. This value is referred to as lbest. The best worth is a worldwide best and is called gbest whenever a molecule accepts all
of humanity as its topological neighbours. The molecular swarm enhancement concept entails altering the speed of each molecule
toward its pbest and lbest values at each time step (for lbest rendition). The amount of time it takes to improve your speed is
proportional to the amount of time it takes to irregular term, with independent arbitrary numbers being created for speed increase
towards pbest and lbest areas.
In the wake of observing the best qualities, the molecule refreshes its speed and positions with following conditions .

w=inertia weight
vi (t) = current velocity
vi (t+1) = modified velocity
xi (t) = current position
xi (t+1) = modified position
c1 & c2 are positive numbers
r1 & r2 are cognitive acceleration and social acceleration components respectively.

IV. RESULTS AND COMPARISON

Comparison between various training algorithm results
TABLE 1: Without using wavelet transformation

S.No Training Method Error MAPE(%) MAE(AUS) Iterations

Function
1 Levenburg Marquardt (trainlm) SSE 5.83 1.58 106
2 Levenburg Marquardt(trainim) MSE 6.20 1.76 116

3 Resilient backpropagation(trainrp) SSE 7.78 2.20 802

4 Resilient backpropagation(trainrp) MSE 6.53 1.84 648
5 Scaled Conjugate SSE 6.35 1.75 432
Gradient(trainseg)
6 Scaled Conjugate MSE 6.61 1.91 288
Gradient(trainseg)
7 Conjugate Gradient with Beale SSE 8.37 2.45 382
Restart(trainegb)
8 Conjugate Gradient with Beale MSE 5.92 1.61 116
Restart(trainegb)
9 Fletcher-Powell Conjugate SSE 7.70 2.25 463
Gradient(traincef)
10 Fletcher-Powell Conjugate MSE 7.73 2.35 830
Gradient(traincef)
11 Polak-Ribière Conjugate SSE 8.88 1.65 407
Gradient(traincap)
12 Polak-Ribière Conjugate MSE 6.91 1.98 447
Gradient(traincap)

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1074
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue II Feb 2022- Available at www.ijraset.com

Fig. 4.1 Comparisons of different training algorithms using SSE error function.

Fig. 4.2 Comparisons of different training algorithms using MSE error function.

Fig. 4.3 Actual and forecast week-ahead NSW electricity price in summer season.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1075
 International Journal for Research in Applied Science & Engineering Technology (IJRASET)
ISSN: 2321-9653; IC Value: 45.98; SJ Impact Factor: 7.538
Volume 10 Issue II Feb 2022- Available at www.ijraset.com

V. CONCLUSIONS
A precise determining power cost is a vital issue of worry for all market members, either for creating offering techniques, or for
settling on speculation choices. In this proposition, a neural organizations model was proposed for value anticipating. The exhibition
of the proposed model is checked utilizing New South Wales (NSW) power value information. Fake neural organizations innovation
is a promising strategy for power value anticipating due to its strong capacity in nonlinear reproduction. In this postulation report, a
multi-facet perceptron fake neural organizations model is built for dayahead power value anticipating in liberated electrical power
market. By the outcomes found in this work it tends to be presumed that the proposed models have produced sensibly great gauge
results. The preparation of the neural organizations with various preparation calculations and mistake capacities gives the agreeable
outcomes. The value information is profoundly unpredictable and non-fixed in nature which should be adjusted for better guaging
results. Wavelet change is utilized for information separating. I have carried out the ANN model and Wavelet-ANN model for
power value determining with various preparation calculation. The blunders can be limited utilizing crossover models which is the
mix of PSO with ANN. To upgrade the loads of neurons molecule swarm enhancement (PSO) procedure has been utilized.

REFERENCES
[1] Murthy, Girish Godekere Panchakshara, Vijayalakshmi Sedidi, Ajaya Kumar Panda, and Badri Narayan Rath. “Forecasting Electricity Prices in Deregulated
Wholesale Spot Electricity Market-A Review.” International Journal of Energy Economics and Policy 4, no. 1 (2013): 32–42.
[2] Prabavathi, M., and R. Gnanadass. “Bidding Strategies for Indian Restructured Power Market.” In Circuits, Power and Computing Technologies (ICCPCT),
2013 International Conference on, 568–73.IEEE, 2013.
[3] Sharma, N. K., P. K. Tiwari, and Y. R. Sood. “Current Status, Policies and Future Perspectives of Indian Power Sector Moving towards Deregulation.” In
Electrical, Electronics and Computer Science (SCEECS), 2012 IEEE Students’ Conference on,1–6.IEEE,2012.
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6184730.
[4] V. P. Gountis and A. G. Bakirtzis, “Bidding strategies for electricity producers in a competitive electricity marketplace,” IEEE Transactions on Power
Systems, vol. 19, no. 1, pp. 356–365, February 2004.
[5] Hamidreza Zareipour, “Price Forecasting and Optimal Operation of Wholesale Customers in a Competitive Electricity Market,” Ph.D. dissertation, Dept.
Electrical and Computer Eng., Univ. of Waterloo, Canada, 2006.
[6] Shahidehpour, M., Yamin, H., and Li, Z. (2002), Market operations in electric Power systems: forecasting, scheduling, and risk management: Wiley.
[7] Weron, R. (2006). Modeling and forecasting electricity loads and prices: a statistical Approach, Chichester: Wiley.
[8] Zareipour, H. (2008), Price-based energy management in competitive electricity markets VDM Verlag Dr.Muller.

©IJRASET: All Rights are Reserved | SJ Impact Factor 7.538 | ISRA Journal Impact Factor 7.894 | 1076