Deep Learning Assisted Online Multi-Step Demand Forecasting of Fused Magnesia Smelting Processes

This paper proposes a multi-step ahead power demand model for fused magnesia smelting processes (FMSP) which combines a linear model and an unknown nonlinear term to predict the electricity demand and its variation tendency for the next 5 steps. The linear model is identified by the multi-output fas...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:2022 4th International Conference on Industrial Artificial Intelligence (IAI) S. 1 - 8
Hauptverfasser: Li, Mingyu, Zhang, Jingwen, Chai, Tianyou
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 24.08.2022
Schlagworte:
Bayesian optimization
Deep learning
Demand forecasting
Long-short term memory
Multi-output fast recursive algorithm
Online multi-step ahead demand forecasting
Power demand
Prediction algorithms
Predictive models
Real-time systems
Smelting
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:This paper proposes a multi-step ahead power demand model for fused magnesia smelting processes (FMSP) which combines a linear model and an unknown nonlinear term to predict the electricity demand and its variation tendency for the next 5 steps. The linear model is identified by the multi-output fast recursive algorithm (MFRA) while the unknown nonlinear term is fitted with a long-short term memory (LSTM) model. The hyperparameters in the LSTM are estimated by the Bayesian optimization (BO) algorithm. Since the sampling period of the power is only 7 seconds, and we have to predict the next 5 steps electricity demand and its tendency within one sampling period, we therefore update parameters of the linear model by the MFRA while parameters of the dense layer of the LSTM are updated by the gradient descent algorithm within the online multi-step demand forecasting framework. The experimental results using the real-time data of a FMSP confirm the effectiveness of the proposed algorithm, achieving up to 52% error reduction in 5-step ahead demand forecasting when compared with other approaches.
DOI:10.1109/IAI55780.2022.9976577