A combination approach based on a novel data clustering method and Bayesian recurrent neural network for day-ahead price forecasting of electricity markets

Abstract

Listen

Translate article

Electricity price forecasting is an ancillary service that plays a key role for market participants in a deregulated structure. Compared to other commodities, electricity price exhibits higher levels of volatility and uncertainty. This imposes more restrictions on the accuracy of the forecast and leads to significant errors. This paper proposes a hybrid electricity price-forecasting framework with a novel time series data mining method to enhance the feature selection. The proposed method includes clustering, preprocessing and training stages. The proposed data clustering method uses both an enhanced game theoretic approach and neural gas in combination with competitive Hebbian learning to provide a better vector quantization. Six strategies are proposed to enable the non-winning neurons to participate in the learning phase and resolve the shortcomings of the original self-organizing map, where the dead neurons are far from the input patterns without having any chance to compete with the winning neurons. The price-load input data are clustered into a proper number of subsets using the proposed data mining method. A novel cluster selection method based on the persistence approach is applied to select the most appropriate cluster as the input to the BRNN. The selected data set is filtered by the harmonic analysis time series, and is time-series processed to provide the proper inputs for training neural networks. Bayesian approach is used to train a recurrent neural network, and forecast the electricity price. The performance of the proposed clustering algorithm is evaluated using different electricity market data. Our results demonstrate the efficiency of the proposed clustering algorithm as compared to K-means, neural gas and self-organizing map clustering methods Our proposed clustering provides 16.7%, 28.6%, and 13% more accurate results than K-means, neural gas, and self-organizing map for the NYISO. CAPITL data. For the NYISO. CENTRL data, the developed clustering outperforms the K-means, neural gas, and self-organizing map by 21.4%, 21.4%, and 8.3%, respectively. The clustering accuracy of the proposed method for NYISO. DUNWOD data is 5.5%, 19%, and 5.5% better than that of the K-means, neural gas, and self-organizing map methods Lastly, for the NYISO. GENESE data, the mean square error value for the proposed clustering is 13.7%, 14.9%, and 12.5% less than that of the K-means, neural gas, and self-organizing map, respectively. The developed forecasting method is also compared with the existing state-of-the-art forecasting algorithms. The comparison results show an improvement in the forecast accuracy of the developed method over other forecasting approaches.