A Multi-Scale Electricity Consumption Prediction Algorithm Based on Time-Frequency Variational Autoencoder

Abstract

Accurate electricity consumption forecasting can be treated as a reliable guidance for power production. However, traditional electricity forecasting models suffer from simultaneously capturing the periodicity and the volatility of sequential electricity consumption data, while the periodicity and the volatility are important for electricity forecasting. In order to effectively model this sequential data and predict electricity consumption accurately, we propose a multi-scale prediction (Long Short Term Memory, LSTM) algorithm based on Time-Frequency Variational Autoencoder (TFVAE-LSTM). The proposed algorithm treats the sequential data as a superposition of data in different frequencies, it defines an encoder in frequency domain to extract frequency features to model the periodicity and volatility, and defines a decoder in time domain to capture the sequential features of data. Based on the extracted Time-Frequency features in a TFVAE, a multi-scale LSTM model is defined to further extract sequential features from different scales to predict electricity consumption. Experiments show the effectiveness of the proposed TFVAE-LSTM for electricity consumption forecasting tasks.

Keywords

• Time-frequency variational autoencoder
• electricity consumption forecasting
• neural network