Household Level Electricity Load Forecasting Using Echo State Network

Abstract

Load forecasting at the household level is challenging because the electricity consumption behavior can be much more variable than those at aggregate levels. The introduction of Advanced Metering Infrastructure (AMI) systems has helped to better forecast the load of an individual household. Since smart meter data is streaming data and there is a need to deal with a massive amount of such data in a real-time fashion, an efficient and fast framework to handle this challenge is required. Deep neural networks like Long Short Term Memory (LSTM) can be used for this purpose but they take a long time to train. In this paper a novel k-nearest meter based Echo State Network (ESN) is proposed and experimental results demonstrate that it is a more suitable candidate for load forecasting, since it is much easier to train and has a great deal of accuracy. The model is compared with other time series models like Persistent (PM) and Vector Autoregression (VAR), as well as deep learning models like multi-layer perceptron (MLP), LSTM and a combination of convolutional neural network (CNN) and LSTM (CNN-LSTM). The results show that the proposed model has a significant improvement over all other models on a dataset spanning 4 months, along with a significant reduction in training time compared primarily to deep learning models.