Summary: The assignment aims at consolidating your knowledge base and developing practical skills to build a fuzzy system for forecasting the electricity price. The task is formulated as a time-series prediction problem for business application, and the goal is to model the behaviour of underlying dynamics of the electricity market. In principle, the merits of such a fuzzy forecasting system can be evaluated by two aspects:

- The number of fuzzy rules in the rule-base and the number of variables used in the antecedent part of the fuzzy rules (the smaller the better);
- System performance in terms of the accuracy (the smaller the better), i.e., the average relative error between your fuzzy system outputs and the actual outputs for both the training data set (learning capability) and the test data set (generalization capability).

The length of the assignment report is about 1200 words.

Problem Description - A Fuzzy System for Forecasting Electricity Price

Develop a fuzzy forecasting system using Matlab Toolbox. The system performs a forecasting task for power marketing price. The data used in this assignment is from the real world (Queensland, Australia), and it has been split up into two parts, i.e., a training dataset which will be used to build your fuzzy forecasting system, and a testing dataset which will be used to evaluate your system performance in terms of generalization capability. The data sets can be downloaded from the Assignment directory in LMS.

Tasks Description

- Remove outliers of the output variable from the datasets (both training and test), and give a list of the outliers; and then rebuild the training and the test datasets;

- Select appropriate values or fuzzy subsets for linguistic variables used in your fuzzy rules;

- List the fuzzy rules that are generated by using statistical analysis (correlation coefficients) with heuristics;

- Implement your fuzzy system using the Matlab Fuzzy Toolbox, where all membership functions involved in your system should be plotted clearly;

- Report your system performance in terms of the average relative error for both training and testing datasets, and analyze the effects of membership functions and defuzzification methods.