کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

DM-LSTM-XGB: A Dual-Memory LSTM with Boosted Error Learning for Robust Smart Building Time-Series Forecasting

نویسندگان :

Ali Sadeghi¹ Mohammed Abdulkareem²

1- دانشگاه صنعتی شیراز 2- University of diyala

کلمات کلیدی :

Smart building،Time series forcasting،DM-LSTM-XGB،Energy Prediction

چکیده :

Reliable energy prediction lies at the very foundation of a smart city, driving optimal energy utilization, demand response, and optimal smart building performance. However, accurate prognosis of energy demand based on multi-modal IoT data has still been challenging: a combination of non-stationarity, ambiguous output from multiple sensors, and persisting prediction error can deteriorate distant-term performance. To overcome these, a new combination forecasting model, named DM-LSTMXGB, combining a dual-memory LSTM network with XGBoost residual learning, is proposed in this article to leverage a unique - yet powerfully robust - forecasting framework that synergistically integrates a dual-memory LSTM network and a residual learning component via XGBoost classification. In this framework, prediction error is considered a form of recurrent memory, where gradients boosted by trees learn to project residual dynamics into the future for pre-activated proactive correction in a repeated loop process in the model structure. The proposed DM-LSTM-XGB model has been carefully tested for performance improvements in a comprehensive real-world dataset for a smart building, encompassing various time series associated with electrical, environmental, and occupancy data, clearly outperforming the performance of both a generic LSTM network and XGBoost in these six performance measures, decreasing error differences by approximately 44.6%, 8.0%, 17.2%, and 2.1%, along with a perceptible improvement in prediction vigor and accuracy, respectively.