Research On Electricity Price Short-term Forecasting

In today's society,energy-related issues continue to affect the city's development and people's lives,and electricity as the 21 st century and human contact one of the most closely related energy,has been a lot of countries through the free trade market to a reasonable distribution.In the free trade market,the price of electricity is not immutable,with a variety of social factors,energy factors such as the occurrence of very frequent fluctuations in the day at different times the price of electricity are not the same,and the price of electricity on the largest The interests of both supply and demand,ease the pressure on the city to promote the rational use of resources has a very significant impact.Because of this,accurate prediction of future electricity prices is particularly important.In this paper,the time-series analysis and multi-feature machine learning of these two different angles,the price of electricity forecasting methods were studied.This paper mainly completed the following work:1)Aiming at the electricity price belongs to the time series domain,this paper used a forecasting method based on pattern sequence similarity,pattern sequence-based forecasting(PSF),this method first obtains the pattern sequence from the clustering stage,and then obtains the pattern sequence based on the sequence pattern similarity.The prediction phase obtains the pattern similarity sequence and the result set through the sliding window,and finally sums the data in the result set in the summation phase.In order to solve the problem of the similarity of historical data,an improved pattern sequence prediction(MPSF)is proposed,which considers the similarity between historical data and current forecast data.Finally,the above two methods are validated on the GEFCom2014-P data set,and the PSF has achieved good prediction results.MPSF has improved the accuracy of PSF.2)A power price forecasting method based on Gradient Boosting Decision Tree(GBDT)is realized for the multi-factor influence of electricity price.The method first excavates the power price characteristics and then uses the feature to supervise the training of the GBDT model.The final trained model has a significant improvement in the prediction accuracy compared to the traditional method on the GEFCom2014-P data set,and provides a set of features that can be used for subsequent research.3)Aiming at the strong nonlinearity of power price data and combining with the feature selection results provided by GBDT,this paper proposes a power price forecasting method based on multi-source data fusion and Long Short-Term Memory(LSTM)This method not only utilizes the memory of LSTM to historical data,but also uses the external factors to influence the power price through the whole connection layer,which provides a new way to solve the problem.This method uses the feature set provided by GBDT to carry out the training of the model,and further improve the accuracy and stability of the forecast on the GEFCom2014-P data set.