Page 52 - C:\Users\am_se\OneDrive - Higher Education Commission\Desktop\FlipBook\

International Conference on

Recent Trends in Environmental Sustainability

ESCON22/CDMP/18
Machine learning based switching model for electricity load forecasting

1
Muhammad Akbar 1,2* , Seyed Safdar Raza , Abdul Mannan 1
1 Department of Electrical Engineering, NFC Institute of Engineering and Technology

Multan, 60000 Pakistan
2 Department of Environmental Sciences, COMSATS University Islamabad, Vehari- Campus,

61100 Pakistan
Correspondence: akbar@cuivehari.edu.pk

Abstract
In vertically integrated power markets of Pakistan, forecasting electricity loads is one of the
most essential tasks for system planning, operation and decision making. The accurate
prediction of service demand is necessary to lay out optimized plans for the generation and
distribution of power. In this context, load forecasting model is pertinent to power system of
Pakistan. Electric load forecasting involves the projection of peak demand levels and overall
energy consumption patterns to support an electricity company’s future system and business
operations. The accurate prediction of both the load magnitude and geographical locations of
electric load over the different time periods becomes. Short and mid-range predictions of power
load allow electricity companies to retain high energy efficiency, reliable operation, and
making optimized plan for power generation and distribution. Power system of Pakistan poses
an eccentric problem, a decade ago generation resources were limited but now it has been
engulfed by circular debt. There is a crisis of capacity payments in energy sector. The present
generation has exceeded to the level that it will be sufficient for next decade but the
transmission and distribution capacity is in dire need of upgradation. Therefore, the market
model has started to shift towards wholesale model instead of present central system.
Competitive Trading Bilateral Contract Market is the first step towards the liberal market.
Keeping in mind the present dynamics, the short-term load forecasting will be of outmost
importance. Therefore, foremost objective of this research is to empower Pakistan’s power
system with accurate load forecasting tool, which will handle the times series as well as other
major variables such as geography, weather and per capita income.

The electricity load forecasting is a complex task due to the nature of load affecting variables.
Dealing with abrupt changes in weather conditions and modeling consumer's usage behaviors
is a challenging task. Machine learning based statistical and artificial intelligence techniques
are widely used. Among these, artificial neural networks (ANN) and support vector machines
(SVM) emerge as competitive modeling approaches. Still, appropriate selection of SVM and
ANN parameters requires proper optimization techniques to avoid slow convergence, local
minima, and over-cutting of models. Another challenge lies in the generalization of these
models; none of the reported models so far, has attained a sturdy stature as a generally
applicable technique. The modern optimization techniques are robust, less time-consuming,
dependable, and provide high quality solutions for parameter selection and model development.
The accuracy of modeling techniques is extremely dependent on quality of historical data.
Since the recording of data in Pakistan power systems was mainly manual. This data contained
abnormalities like missing values, outliers, and duplication of records. Observing all the
aforementioned problems, we got motivation to devise such a model that can perform well on

Department of Environmental Sciences, COMSATS University Islamabad, Vehari Campus

19