Decomposition-Residuals Neural Networks: Hybrid system identification applied to electricity forecasting

COVID-19 causes issues in energy demand profile forecasting. Day-ahead energy forecasting systems struggle to provide accurate demand predictions. Decomposition-Residuals Deep Neural Networks (DR-DNN) is a two-layer hybrid architecture with: a decomposition and a nonlinear layer. Based on statistical tests, the decomposition layer applies robust signal decomposition and low-order system identification of input data streams into: trend, seasonal and residuals signals. Utilizing calendar information, temporal signals are added: sinusoidal day/night cycles, weekend/weekday, etc. The nonlinear layer can learn unknown nonlinear patterns using typical DNN models like Recurrent Neural Networks (RNN) and Convolutional Neural Networks (CNN). DR-DNN outperformed baseline, linear, and plain DNN models on next-day electricity forecasts within the COVID-19 era (September 2020-January 2021).

新冠疫情给能源需求负荷曲线预测带来了显著挑战，日前能源预测系统难以生成精准的需求预测结果。分解残差深度神经网络（Decomposition-Residuals Deep Neural Networks，DR-DNN）是一种双层混合架构，包含分解层与非线性层。基于统计检验，分解层可对输入数据流实施鲁棒信号分解与低阶系统辨识，将其拆解为趋势信号、季节信号与残差信号三类。结合日历信息，还可补充时序特征信号，例如正弦昼夜周期、周末/工作日标识等。非线性层可借助典型深度神经网络（Deep Neural Networks，DNN）模型，如循环神经网络（Recurrent Neural Networks，RNN）与卷积神经网络（Convolutional Neural Networks，CNN），学习未知的非线性模式。在2020年9月至2021年1月的新冠疫情期间，DR-DNN在次日电力预测任务中的表现优于基准模型、线性模型与普通深度神经网络模型。