This thesis reports findings from a number of modern machine learning techniques applied to electricity market price forecasting. The techniques evaluated were Support Vector Machines, Boosting, Bayesian networks, neural networks and a weekly average method. All techniques were evaluated on seven day into the future forecasting of the Regional Reference (pool) Prices (RRP) for the New South Wales (NSW) region of the Australian National Electricity Market (NEM). Due to highly volatile and non-repetitive nature of the NSW RRP, all complex machine learning methods provided inferior accuracy forecasts compared to a weekly average method. The weekly average method was computationally less expensive and more transparent to the user than any of the machine learning techniques. The Support Vector Machine (SVM) was chosen for its novel application to electricity price forecasting because it is considered to be the next generation to neural networks. The structured SVM training algorithm proved more consistent and reliable than the neural network algorithm. Bayesian networks offer the adaptability of a neural network with the advantage of providing a price forecast with confidence intervals for each half-hour determined from the actual data. The SVM and Bayesian techniques were found to provide acceptable forecasts for NSW demand. An investigation of international electricity markets found that each market was unique with different market structures, regulations, network topologies and ownership regimes. Price forecasting techniques and results cannot be universally applied without careful consideration of local conditions. For instance, price data for the Spanish and Californian electricity markets were investigated and found to have significantly lower price volatility than the NSW region of the NEM. An extensive examination of the NSW RRP showed that the price exhibited no consistent long-term trend. A stationary data set could not be extracted from the price data. Thus, making forecasting unsuited to techniques using large historical data sets. The strongest pattern found for NSW prices was the weekly cycle, so a weekly average method was developed to utilise this weekly cycle. Over 25 weeks of NSW RRP from February to July 2002, the seven day into the future price forecast mean absolute error (MAE) for the SVM technique was 27.8%. The weekly average method was more accurate with an MAE of 20.6% and with a simple linear price adjustment for demand, the error was reduced to 18.1%. The price spikes and uneven distribution of prices were unsuitable for the Boosting or Bayesian network techniques.
Identifer | oai:union.ndltd.org:ADTP/279089 |
Creators | Sansom, Damien |
Source Sets | Australiasian Digital Theses Program |
Detected Language | English |
Page generated in 0.0015 seconds