Scholarlink

<p>Accurate energy consumption forecasting is crucial for efficient electrical power management in smart environments like smart buildings, enabling significant cost savings and optimized operational strategies. This paper introduces a novel hybrid deep learning-based approach to predict total active energy consumption. Our method employs robust data preprocessing to ensure the model receives high-quality input data. Subsequently, we decompose the original signal into a residue and multiple Intrinsic Mode Functions (IMF) using Ensemble Empirical Mode Decomposition (EEMD). The next step involves the application of an Encoder–Decoder architecture with Long Short-Term Memory (ED-LSTM) layers to capture historical dependencies and generate future predictions. An Efficient Channel Attention (ECA) mechanism is incorporated to enhance feature selection by adaptively reweighting the most informative features. The proposed EEMD-ED-LSTM-ECA model is applied to each IMF component and the residue individually. The final prediction is obtained by aggregating the partial predictions from all components. To evaluate the model’s effectiveness, a residential smart building dataset is employed. The EEMD-ED-LSTM-ECA model is designed to generate forecasts for various time horizons (1-, 4-, and 24-h) with a 1-h resolution level. To assess its performance, we compared the performance of the proposed EEMD-ED-LSTM-ECA model with those of EEMD-ED-LSTM, ED-LSTM, and other state-of-the-art models through the following metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R<sup>2</sup>). The developed methodology outperformed the compared models across all forecasting horizons. For the 1-h forecast, it achieved an MAE of 0.228 kW, an RMSE of 0.335 kW, and an R<sup>2</sup> of 0.715. For 4-h predictions, the MAE, RMSE, and R<sup>2</sup> were 0.085 kW, 0.127 kW, and 0.968, respectively. Over a 24-h horizon, it reached an MAE of 0.113 kW, an RMSE of 0.164 kW, and an R<sup>2</sup> of 0.944. These results demonstrate the model’s high accuracy and robustness for short- and long-term forecasting. These results highlight the model’s potential to significantly improve energy management in smart buildings by providing highly accurate predictions for both short-term operational adjustments and long-term strategic planning.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Smart Building Energy Forecasting Using Hybrid Approach

  • 期刊论文

  • 发表日期：

    2025年8月30日

查看全文
  • Chaymae Makri,
  • Said Guedira,
  • Imad El Harraki,
  • Soumia El Hani

摘要

Accurate energy consumption forecasting is crucial for efficient electrical power management in smart environments like smart buildings, enabling significant cost savings and optimized operational strategies. This paper introduces a novel hybrid deep learning-based approach to predict total active energy consumption. Our method employs robust data preprocessing to ensure the model receives high-quality input data. Subsequently, we decompose the original signal into a residue and multiple Intrinsic Mode Functions (IMF) using Ensemble Empirical Mode Decomposition (EEMD). The next step involves the application of an Encoder–Decoder architecture with Long Short-Term Memory (ED-LSTM) layers to capture historical dependencies and generate future predictions. An Efficient Channel Attention (ECA) mechanism is incorporated to enhance feature selection by adaptively reweighting the most informative features. The proposed EEMD-ED-LSTM-ECA model is applied to each IMF component and the residue individually. The final prediction is obtained by aggregating the partial predictions from all components. To evaluate the model’s effectiveness, a residential smart building dataset is employed. The EEMD-ED-LSTM-ECA model is designed to generate forecasts for various time horizons (1-, 4-, and 24-h) with a 1-h resolution level. To assess its performance, we compared the performance of the proposed EEMD-ED-LSTM-ECA model with those of EEMD-ED-LSTM, ED-LSTM, and other state-of-the-art models through the following metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R2). The developed methodology outperformed the compared models across all forecasting horizons. For the 1-h forecast, it achieved an MAE of 0.228 kW, an RMSE of 0.335 kW, and an R2 of 0.715. For 4-h predictions, the MAE, RMSE, and R2 were 0.085 kW, 0.127 kW, and 0.968, respectively. Over a 24-h horizon, it reached an MAE of 0.113 kW, an RMSE of 0.164 kW, and an R2 of 0.944. These results demonstrate the model’s high accuracy and robustness for short- and long-term forecasting. These results highlight the model’s potential to significantly improve energy management in smart buildings by providing highly accurate predictions for both short-term operational adjustments and long-term strategic planning.