Abstract: Under the big data background of smart grid popularization, accurate analysis and forecasting of power data is of great guiding significance to the planning of power grids and the management decisions of economic departments. However, most of the models are only studied on a single time scale. Aiming at this problem, a recurrent neural network prediction model based on time series decomposition back propagation algorithm is proposed. Through the statistical analysis of the real consumer electricity consumption data and external factors, the residential electricity consumption characteristics and behavior rules are analyzed in depth. According to the characteristics of the data and external factors such as weather, holidays and other factors affecting the user’s electricity behavior, a forecasting model is established to predict the electricity consumption of users in the future. In addition, taking into account the temporal characteristics of household electricity consumption data showing different variation rules at different time scales, the forecasting model is established by time series decomposition to model the periodicity and trend of user’s electricity consumption behavior and to achieve the goal by weighted fusion training effect, with a certain degree of synergy, improving prediction accuracy.

Key words: prediction model, back propagation algorithm, recurrent neural network, time series decomposition, power data

摘要： 在智能电网普及的大数据背景下，对电力数据进行精准的分析和预测对电网规划和经济部门的管理决策具有重要的指导意义，但大多数模型都只是在单一的时间尺度上进行研究。针对这一问题提出一种基于时序分解的后向传播算法的循环神经网络预测模型。通过对真实的居民用电消费数据以及外部因素数据统计处理，深入地分析了居民用电特点以及行为规律，并根据其数据的特征以及天气、节假日等外部因素对用户用电行为的影响建立预测模型，对用户未来时段的用电量进行预测。此外，考虑到居民用电消费数据的时序特征在不同时间尺度呈现不同的变化规律，通过时序分解建立预测模型来对用户用电行为的周期性和趋势性进行建模，并通过加权融合达到一起训练的效果，具有一定的协同性，提升预测精度。

关键词: 预测模型, 后向传播算法, 循环神经网络, 时序分解, 电力数据