TL;DR
This paper introduces a novel approach for improving short-term load forecasting in integrated energy systems (IES) by combining outlier detection with a sophisticated multi-task learning model. By effectively cleaning and enriching the data before feeding it into a TCN-MMoL network, the method significantly enhances prediction accuracy, crucial for stable and economical IES operation.
What Problem Does This Paper Address?
Accurate short-term load forecasting is critical for the efficient management and scheduling of Integrated Energy Systems (IES). These systems, which combine various energy sources like electricity, heating, and cooling, are becoming increasingly complex and dynamic. Traditional forecasting methods often struggle with the "unconventional data" that arises from system anomalies, extreme events, or data transmission errors. This unconventional data can severely degrade the performance of forecasting models, impacting the stability and economic efficiency of the IES. The paper aims to address this by developing a robust method that can effectively handle and correct for these data anomalies before making predictions.
Methods
The proposed method involves a multi-stage approach:
- Data Preprocessing:
- Isolation Forest (iForest): This unsupervised anomaly detection algorithm is used to identify outliers in high-dimensional data matrices constructed using a sliding window technique. iForest is chosen for its efficiency in handling large datasets and its ability to detect anomalies without prior labeling.
- Dynamic Orbit Algorithm: This novel approach is used to repair the identified abnormal data. It defines a "channel" (upper, middle, and lower orbits) based on historical data patterns. Data points falling outside this channel are flagged as anomalies and then corrected using a baseline value derived from the normal range. This method is specifically designed to handle the volatility and randomness of IES load data.
- Correlation Analysis (AR and MIC): * Autoregressive (AR) model: Used to analyze the temporal characteristics of the load data and determine parameters for the dynamic orbit method.
- Maximal Information Coefficient (MIC): Employed to uncover complex linear and nonlinear relationships between various loads (cooling, heating, electric) and weather data. This helps construct a high-dimensional feature matrix that captures strong correlations.
- Multi-Load Prediction Model:
- TCN-MMoL (Temporal Convolutional Network - Multi-gate Mixture-of-Experts LSTM): This multi-task learning framework is used for the final prediction.
- TCN: Extracts temporal and spatiotemporal features from the preprocessed high-dimensional data.
- LSTM: Further refines the feature matrix, leveraging its memory capability to capture temporal dependencies.
- MMoL (Multi-gate Mixture-of-Experts): This architecture, inspired by Google's MMoE, allows for specialized "expert" networks to process different aspects of the data, with a gating mechanism dynamically weighting their contributions for each specific load (cooling, heating, electric). This allows for better handling of diverse correlations between different load types.
Datasets
The study utilizes load data from the integrated energy system of the Tempe campus of Arizona State University. This dataset includes cooling load, heating load, electric load, and meteorological data. The data spans from January 2022 to May 2023 with a sampling granularity of 15 minutes. The dataset was divided into 75% for training, 15% for verification, and 10% for testing.
Key Insights and Ideas
- Anomaly Detection is Crucial: The paper strongly emphasizes that raw, uncleaned data, especially with anomalies, significantly hinders the accuracy of IES load forecasting.
- Dynamic Orbit for Robust Repair: The proposed "dynamic orbit" method offers a more sophisticated way to repair anomalies in time-series data compared to simple thresholding, by considering the historical "channel" of normal behavior.
- MIC for Capturing Complex Relationships: The Maximal Information Coefficient proves effective in identifying not just linear but also complex nonlinear dependencies between different energy loads and environmental factors, which is vital for IES.
- TCN-MMoL for Specialized Multi-Task Learning: The TCN-MMoL architecture, with its mixture-of-experts approach, allows for more nuanced learning of inter-load dependencies than traditional shared-layer multi-task models.
Key Results
The proposed IDTMMoL model demonstrated significant improvements in prediction accuracy compared to baseline models:
- Cooling Load: MAPE reduced by 2.57% and MAE was also reduced compared to a model using only iForest for preprocessing (ITMMoL).
- Electric Load: MAPE reduced by 0.49%.
- Overall: The IDTMMoL model achieved the lowest MAPE and MAE across all three load types (cooling, heating, and electric) in the test set, outperforming both the ITMMoL (iForest + TCN-MMoL) and IDCLSTM (iForest + Dynamic Orbit + CNN-LSTM) models. For instance, the IDTMMoL model achieved a MAPE of 3.80% for cooling load, significantly lower than the IDCLSTM's 6.98%.
Limitations and Future Work
- Limitations: The paper focuses on a specific integrated energy system. The generalizability of the proposed method to other IES configurations or different types of energy systems might require further validation. While the dynamic orbit method is presented as novel, its theoretical underpinnings and robustness across a wider range of anomaly types could be further explored.
- Future Work: The authors suggest extending their research to forecasting applications in wind power, photovoltaic power, and electricity prices. Further exploration into more application scenarios for their developed methods is also proposed.
Summary updated