基于周期性建模的时间序列预测方法及电价预测研究

时间序列数据广泛存在于人类的生产生活中, 通常具有复杂的非线性动态和一定的周期性. 与传统的时间序列分析方法相比, 基于深度学习的方法更能捕捉数据的深层特性, 对具有复杂非线性的时间序列有较好的建模效果. 为了在神经网络中显式地建模时间序列数据的周期性和趋势性, 本文在循环神经网络的基础上引入了周期损失和趋势损失, 建立了基于周期性建模和多任务学习的时间序列预测模型. 将模型应用到欧洲能源交易所法国市场的能源市场价格预测中, 结果表明周期损失和趋势损失能够提高神经网络的泛化能力, 并提高预测时间序列趋势的精度.

Time Series Forecasting Based on Seasonality Modeling and Its Application to Electricity Price Forecasting

Time series data exist widely in human production and life. The real time series data often contain complex nonlinear dynamics and seasonality. Compared with traditional time series analysis methods, deep learning based methods have good modeling effect for the time series with complex nonlinearities but fail to model the seasonality and trend of time series. In order to model the seasonality and trending explicitly in neural networks, this paper introduces seasonal loss and trend loss into recurrent neural networks (RNNs), establishing the time series prediction model based on seasonality modeling and multi-task learning. The suggested method is then applied to the electricity price forecasting on EPEX (European Power Exchange) France market. The experiment results show that seasonal loss and trend loss can improve the generation ability of neural networks and the performance of sequence trend forecasting.