TL;DR
This paper introduces MEC-TLL, a novel framework for forecasting electric energy consumption across multiple profiles in smart buildings. By combining k-means clustering with transfer learning for Long Short-Term Memory (LSTM) models, MEC-TLL significantly reduces computational time while maintaining high predictive accuracy, addressing a key bottleneck in smart building energy management.
What Problem Does This Paper Address?
Accurately forecasting electric energy consumption is crucial for efficient energy management in smart buildings, enabling better resource allocation and anomaly detection. Existing approaches often focus on predicting consumption for a single profile (e.g., one apartment) at a time. However, a smart building comprises numerous individual profiles, and retraining predictive models for each one is computationally expensive and time-consuming. This creates a significant challenge for real-time, large-scale energy management systems. The paper aims to bridge this gap by developing a method that can efficiently forecast energy consumption for multiple profiles simultaneously.
Methods
The core of the proposed MEC-TLL framework lies in its innovative combination of techniques:
- K-Means Clustering: The daily load demand data from multiple profiles is first clustered using the k-means algorithm. This groups similar consumption patterns together, allowing for more efficient processing. Silhouette analysis is employed to determine the optimal number of clusters for each dataset.
- Transfer Learning: Instead of training an LSTM model from scratch for every single profile, the framework leverages transfer learning. A "base" LSTM model is trained on the centroid (representative time series) of each cluster. This pre-trained model then serves as a starting point for training models for all other profiles within that same cluster. This "transfers" learned knowledge, significantly reducing the training time for individual profiles.
- Long Short-Term Memory (LSTM) Networks: LSTMs are used as the underlying predictive model due to their proven effectiveness in handling time series data and capturing long-term dependencies, which are characteristic of energy consumption patterns.
The overall workflow involves: data preprocessing (transformation, noise treatment), clustering the training data, training a base LSTM model for each cluster centroid, transferring this knowledge to train models for all profiles within each cluster, and finally, using these trained models for prediction on the testing data.
Datasets
The study utilizes two real-world datasets collected from smart buildings in South Korea. These datasets contain 15-minute demand loads for multiple apartments (profiles) over three years (January 2016 to December 2018).
- B1 Dataset: Consists of 96 profiles.
- B2 Dataset: Consists of 91 profiles.
The data was converted to daily demand loads for the experiments. The first two years of data were used for training, and the last year for testing.
Key Insights and Ideas
- Clustering as a Precursor to Transfer Learning: The paper highlights that simply applying transfer learning might not be optimal if profiles are highly diverse. Clustering profiles with similar energy consumption patterns first allows transfer learning to be applied more effectively within homogeneous groups, maximizing knowledge transfer.
- Computational Efficiency is Key: The primary motivation is to drastically reduce the computational burden of forecasting for many profiles. The results strongly suggest that this goal is achieved, making large-scale smart building energy management more feasible.
- Transfer Learning Benefits from Structure: The comparison between TL-LSTM (transfer learning without clustering) and MEC-TLL (transfer learning with clustering) demonstrates that the cluster-based strategy provides an advantage, leading to better predictive performance in addition to computational savings.
Key Results
The experimental results demonstrate the effectiveness of the MEC-TLL framework:
- Computational Time Reduction:
- On the B1 dataset, MEC-TLL achieved an average computational time of 15.7 seconds per profile, compared to 95.4 seconds for TML-LSTM (traditional LSTM per profile) and 17.2 seconds for TL-LSTM. This represents a significant speedup, with the proposed method taking around 25 minutes for the entire dataset, compared to 2.7 hours for TML-LSTM.
- On the B2 dataset, MEC-TLL averaged 15.4 seconds, compared to 48.2 seconds for TML-LSTM and 16.8 seconds for TL-LSTM.
- Predictive Performance:
- For the B1 dataset, MEC-TLL achieved RMSE of 1.142, MAE of 0.670, and MAPE of 34.32. These metrics were comparable to or slightly better than TML-LSTM and significantly better than TL-LSTM.
- For the B2 dataset, MEC-TLL achieved RMSE of 1.368, MAE of 0.891, and MAPE of 41.07, again showing competitive performance.
- Optimal Clusters: Silhouette analysis determined optimal cluster numbers of 11 for the B1 dataset and 6 for the B2 dataset.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the predictive performance of MEC-TLL and TML-LSTM are "almost the same," suggesting that while computational time is drastically reduced, there might be a slight trade-off in predictive accuracy compared to training each profile individually from scratch (though MEC-TLL still outperforms TL-LSTM).
- The TL-LSTM approach (transfer learning without clustering) showed "poor predictability" compared to TML-LSTM and MEC-TLL, indicating the importance of the clustering step.
- Future Work: The authors propose several avenues for future research:
- Enhancing performance using other modern techniques like Bi-directional LSTMs, Discrete Wavelet Transform (DWT), and Discrete Fourier Transform (DFT) for feature extraction.
- Adapting the proposed framework for real-time environments to support industrial applications.
Summary updated