Murdoch University Research Repository

Welcome to the Murdoch University Research Repository

The Murdoch University Research Repository is an open access digital collection of research
created by Murdoch University staff, researchers and postgraduate students.

Learn more

Introducing system-based spatial electricity load forecasting

Kardehi Moghaddam, Mehdi (2016) Introducing system-based spatial electricity load forecasting. PhD thesis, Murdoch University.

PDF - Whole Thesis
Download (8MB) | Preview


The main motivation of this research is to help reduce the Green House Gases (GHG) emissions of the electricity sector, and counteract the effects on nature and people. Traditional methods of power planning are not optimised to achieve this, and only consider Capital Expenditure (Capex) and Operational Expenditure (Opex) reduction as their main objectives. Minimising GHG emissions is now an additional objective of power planning. One way of achieving this is by optimising the distance of generators to the loads to reduce the transmission losses, and also by harnessing the available regional sources of renewable energies and increasing their integration in the network. Efficient load forecasting methods, capable of describing the regional behaviours of the electricity consumption are developed in this research, and can provide priceless input to electricity planners. Such forecasting methods, known as spatial forecasting, can be used to extract short-term and medium-term information of the electricity consumption of different regions. This work also provides tools for making decisions about the most accurate way of pre-processing consumption data and choosing the most efficient forecasting procedure.

Chapter 1 talks about emissions of GHGs and their adverse effect on the nature. It introduces electricity sector as one of the major contributors of human made GHG emissions. It then describes the components of electrical power network and the planning of it. Finally the chapter concludes that an efficient spatial load forecasting method is required to help with spatial planning of power networks. The spatial planning can include more regional components like proximity of generation components to consumers, or the levels of harnessed renewable energy in each area. In such an approach, GHG reduction can be also considered along with Capex and Opex minimisation to plan the future of power networks.

Chapter 2 provides definitions on power network components and the load forecasting methods. It starts with definition of power systems and explanation on how electrical energy is superior to all other forms of energy from end user point of view. Electricity generation systems and the sources of energy to produce electricity are described next. Typical generation unit sizes in MW, continuity of the supply, and also its predictability are summarised in a table at the end of this section. Thereafter, transmission lines and distribution systems are described, as other component of electrical power networks. Importance of having an accurate forecast of electricity demand and the common ways to do it are presented next. At the end of this chapter, the deficiencies of current forecasting methods are highlighted and one major goal is defined for this work. It is to overcome the deficiencies of individual forecasting methods by combining them and using them only where it performs efficient. It also mentions that the work is going to closely look at the behaviour of input data to the forecasting method to seek better methods for preparing them.

Chapter 3 describes South West Interconnected System (SWIS) as the case study for this work. The reasons for selecting SWIS as the case study are mentioned, followed by a quick history of it and how it has been expanded over the last hundred years. To be able to complete spatial forecasting, the area under study needs to be divided into regions. SWIS is then divided into eight regions for this purpose. A visual presentation of the eight regions on the map is presented at the end of this chapter for more clarity.

Chapter 4 performs a short forecasting method on one of the SWIS regions. The selected region is called Metro East. Metro East region is mainly composed of residential consumers. Unlike commercial and industrial consumers, the residential ones are not following a working schedule. That's why it makes them to behave differently and more randomly comparing to the other two. This means more complicated demand to forecast. This is the main reason that Metro East is selected to be studied on this chapter. One of the main components of this chapter is to introduce the methods that have been used for pre-processing of input data. The pre-processing stages include data resolution adjustment, replacement of missing data, removing outliers, clustering and signal reconstruction. A well pre-processed set of data is critical component of any forecasting strategy. The second component of chapter 4 is to generate one day ahead and seven day ahead forecasts of Metro East electricity consumption, using three different training methods. The forecasted results are comparable to other studies done on short term load forecasting. However the author questions the accuracy of classic approach of load forecasting. Classic approach is basically what have been done in the field of load forecasting for decades, which is very similar to the works done in chapter 4. In classic approach, a method gets tested on a case study with an acceptable level of accuracy. Then that method gets introduced as a very accurate tool to be applied on demand forecasting purposes. This work is showing that such accurate method cannot be accurate at all when being applied to other different case studies. Future chapters study this in further details, and come up with some guidelines on how to have accurate load forecast based on the nature of the case study in hand.

Chapter 5 applies the methods of load forecasting developed in chapter 4 onto eight different case studies. By doing this, it can be seen that there is no single method of forecasting that can be accurate for all case studies out there. Temperature sensitivity and distribution of the load data of all the regions is closely studied for fifteen years of data. A load type determination criterion is presented in Table 5. By using this table, and preparing Rayleigh, Generalised Pareto, and Generalised Extreme Value distributions of the load data under study, anyone will be able to say whether their load under study is mainly commercial, residential or industrial. The outdoor temperature is one of the main inputs of short term electricity forecasting. Same chapter shows that residential loads are having a greater temperature sensitivity comparing to the other two. The results of one day and seven day ahead forecasts of the eight regions are presented at the end of chapter 5, using two methods of neural networks and decision trees. The results suggest that the two methods need to be used alternatively based on the characteristics of the case study and ambient temperature to achieve the best result.

Chapter 6 explains the system based medium term load forecasting. The approach to medium term forecasting is completely different to the one developed for the short term one. Two main differences between Short-Term Load Forecasting (STLF) and Medium-Term Load Forecasting (MTLF) are the availability of weather data and the forecasting objectives. Because of the nature of the weather, temperature forecasts of a year ahead are completely impossible. Also in medium term load forecasting the focus of planners is mainly on peak load and energy consumption forecasts. The forecasting method presented in this chapter is achieved by superimposing annual trend, annual seasonality and forecasted residuals by neural networks and decision trees. Similar to chapter 5, the forecasting strategy is applied to eight different case studies for comparison. It is concluded that based on the case under study, the accuracy of the methods changes. It also provides some advices on the best practices to perform medium load forecasting, considering the characteristics of the load. For instance, it conclude that for industrial regions regression trees performs better than neural network based methods. The same applies to CBD region where commercial load dominates. For some residential areas neural networks behave better. This is because of higher nonlinearity of residential load.

The major contributions of this work can be summarised as below:
- The topic of the study, i.e. spatial load forecasting and the potential of using it in efficient power planning, is relatively a new topic in the electricity market literature. Moreover, many of the known spatial load forecasting methods have not yet been widely used because of the size, variety, and availability of the data required. The methodology proposed in this study can successfully be applied to spatial forecasting.
- While conventional methods are useful for short-term predictions with acceptable accuracy, they fail when medium-to-long term load forecasting is dealt with. The methodology conceived and implemented in this thesis is significantly better than those known as state-of-the-art and can give very satisfactory results for medium-term predictions.

- The load analysis criterion, particularly using Q-Q (Quantile vs. Quantile) plots is a unique and original finding of this work. While Q-Q plots are largely used in traditional statistics to compare two samples of data, it has never been applied before for electricity load forecasting purposes. Based on its definition and use, an electricity planner can understand which part of the load is the dominating factor (i.e. whether it is residential, commercial or industrial). And then, based on this, he/she can decide how to go ahead with choosing the most effective forecasting method. Based on this, the thesis provides a very useful criterion for decision making in the energy market.

- One of the major findings of the thesis is that there is no one optimum way of forecasting electricity load in different scenarios. The results presented in the thesis have shown that a method that can accurately forecast the load on a system (3% error for a year ahead) can perform completely different in forecasting another system (observed errors of around 14%). This study demonstrates that a method which is claimed to have a given accuracy can be considerably inaccurate when applied on a different case study.

- Using an ambient temperature-based criterion (i.e. the average maximum temperature of the month) to choose the correct forecasting method is another major finding of the study. In fact, the author has demonstrated that for a temperature sensitive load, different forecasting methods should be used and then combined to get the most accurate result.

Item Type: Thesis (PhD)
Murdoch Affiliation(s): School of Engineering and Information Technology
Supervisor(s): Bahri, Parisa
Item Control Page Item Control Page


Downloads per month over past year