Multi-variable Adversarial Time-Series Forecast Model

TL;DR

This paper introduces a multi-variable adversarial LSTM framework for power system forecasting, improving accuracy by jointly modeling multiple electric variables and their relations, with promising results on industrial enterprise data.

Contribution

It presents a novel adversarial training approach for multi-variable time-series forecasting in power systems, enhancing accuracy and variable relation modeling.

Findings

Improved electricity consumption prediction accuracy.

Effective joint modeling of multiple power system variables.

Successful application to real industrial enterprise data.

Abstract

Short-term industrial enterprises power system forecasting is an important issue for both load control and machine protection. Scientists focus on load forecasting but ignore other valuable electric-meters which should provide guidance of power system protection. We propose a new framework, multi-variable adversarial time-series forecasting model, which regularizes Long Short-term Memory (LSTM) models via an adversarial process. The novel model forecasts all variables (may in different type, such as continue variables, category variables, etc.) in power system at the same time and helps trade-off process between forecasting accuracy of single variable and variable-variable relations. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples. The predict results of electricity consumption of industrial enterprises by multi-variable adversarial time-series forecasting model show that the proposed approach is able to achieve better prediction accuracy. We also applied this model to real industrial enterprises power system data we gathered from several large industrial enterprises via advanced power monitors, and got impressed forecasting results.