Font Size: a A A

Study Of Time Series Forecasting Methods And Electricity Price Forecasting Based On Deep Learning

Posted on:2020-11-01Degree:MasterType:Thesis
Country:ChinaCandidate:R C XuFull Text:PDF
GTID:2518306185499684Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
The time series is an important data form.It is of great practical significance to perform effective analysis and accurate forecasting of time series.Time series data usually has long-term complex nonlinearity and high volatility.Traditional time series analysis methods are not effective enough and faced with difficulty in model and parameter selection,poor flexibility and applicability problems.Deep learning is very appropriate for time series modeling due to its strong nonlinearity,high precision,and suitability for large samples.Seasonality is the remarkable feature of time series data,and theoretically seasonal features are very helpful for tasks like time series forecasting.For time series containing seasonality,this paper chooses time series seasonal decomposition as the starting point and develops a series of time series prediction algorithms based on deep learning.The proposed algorithms are verified on electricity price data of the European Power Exchange in France.The main research contents include:(1)The principle and defects of commonly used time series seasonal decomposition algorithms are analyzed.In order to improve the defects,a time series periodic decomposition model based on gated multi-layers recurrent neural network and the corresponding constraint loss functions are proposed.The decomposition results show that the proposed model has the advantages of flexibility,high efficiency,good real-time performance and strong anti-disturbance capability,which is more advantageous than the traditional algorithm.(2)This paper proposes a seasonal decompose based model,combining seasonal decomposition model with the time series forecasting model to help the model learn long-term dependence.The model effectively improves the accuracy of time series forecasting.For the long-term forecasting of time series,two possible models with seasonal decomposition are proposed,which have a stable long-term forecasting effect.(3)Analyze the uncertainty of time series prediction and point out the necessity of prediction intervals estimation.The applicability of the Lower Upper Bound Estimation(LUBE)method and the improved LUBE method to time series prediction intervals estimation is analyzed.This paper further proposes the RNN-based prediction intervals estimation model and its detailed training process.On this basis,model trained with the data that eliminates the seasonal sequence can significantly improve the estimation effect of the prediction intervals.
Keywords/Search Tags:Time Series Seasonal Decomposition, Time Series Forecasting, Prediction Intervals, Deep Learning, Recurrent Neural Networks, Energy Market
PDF Full Text Request
Related items